CHES 2017 Taipei, Taiwan

CHES 2017 was held September 25th - 28th in Taipei, Taiwan. This being my first trip to CHES, I was glad to see a mix of academics and people in industry whom had ties with cryptographic hardware on embedded systems.

Although I have a limited frame of reference, I feel the standard of the conference was quite high - the presenters all knew what they were talking about in great detail, and were able to accurately describe the contribution they had made to their respective fields.

My favourite talks were in the 'Side-Channel Analysis' and the 'Emerging Attacks' sessions, as the talks in these two sessions in particular were engaging and close to the work I have been doing during my PhD.

However, my obligatory post-conference blog post will be on 'Sliding right into disaster: Left-to-right sliding windows leak', a joint work by Daniel J. Bernstein, Joachim Breitner, Daniel Genkin, Leon Groot Bruinderink, Nadia Heninger, Tanja Lange, Christine van Vredendaal, and Yuval Yarom (I wasn't aware so many people could work on one paper at the same time!).

The contribution of the paper was showing that although the Right-to-Left sliding window didn't provide leak a great deal of information, the Left-to-Right sliding window provided just enough to recover the full key (in some cases).

For a brief recap, RSA uses modular exponentiation, and in many implementations the 'sliding window' method is used for efficiency. This can be done either Left-to-Right or Right-to-Left, and although they are very similar, they have very slight differences: the Right-to-Left method tends to be easier to program, uses the same number of multiplications as Left-to-Right, but requires a bit more storage. Both are used in practice: the paper shows that the Libgcrypt crypto library uses the Left-to-Right method (and hence they provide an attack against this implementation).

One way to think about it is that if you want to compute x^25, you would convert the exponent 25 into binary, manipulate this bitstring in some way (depending on whether you are going Left-to-Right or Right-to-Left, and also on the size of your window w), and then parse the bitstring: for every non-zero bit, perform a multiply; for every zero bit, perform a square (or something to that effect)

In this manipulated bitstring in the Right-to-Left method, due to the way the bitstring is created, we are guaranteed to have w - 1 zero bits after a non-zero bit. From a leakage point of view, this doesn't provide much information.

However, in the Left-to-Right method, two non-zero bits can be as close as adjacent. This allows us to infer certain details about the bitstring by applying certain rules to what we know (the leakage), and in some cases, working out the value of the key.

If we are able to recover >50% of the key bits this way, we can implement an efficient Heninger-Shacham attack to recover the remaining bits.

The paper was presented by Leon Groot Bruinderink, and he explained it in such a way that I found it clear to understand how the attack works, and how one would prevent against this kind of attack (not using Left-to-Right would be a start). They also contacted Libgcrypt with details of the attack, and it has been fixed in version 1.7.8.

Aside from the papers, CHES has been pretty amazing: the venue was a 5 star hotel in the centre of Taipei, the food was absolutely incredible (even the banquet, which introduced me to the wonders of sea cucumber), and the excursion to the Taipei Palace Museum was exceptionally educational (which as we all know is the best kind of fun).

I would definitely recommend CHES to anyone interested in the more practical side of cryptography, although if it ever takes place in Taiwan again, I strongly suggest you Youtube how to use chopsticks. Unfortunately I never learnt, and after a humiliating trip to the ShiLin Night Market, am now featured on several locals' phones in a video named 'The Tourist who couldn't eat Beef Noodle Soup'.

Crypto 2017 - How Microsoft Wants to Fix the Internet (Spoiler: Without Blockchains)

In the second invited talk at Crypto, Cédric Fournet from Microsoft Research presented the recent efforts of Project Everest (Everest VERified End-to-end Secure Transport), which seems an attempt to fix implementing TLS once and for all. Appropriately for such a gigantic task, more than a dozen researchers on three continents (and the UK) work on making it verifiable and efficient at the same time.

As with every self-respecting talk in the area, it started with the disaster porn that is the history of TLS (our lifes depend on it, but it's complicated, there have been 20 years of failures, insert your favourite attack here). However, the Crypto audience hardly needs preaching to (just a reminder that the work isn't done with a proof sketch; performance and error handling also matters), so the talk swiftly moved on to the proposed solutions.

The story starts in 2013 with miTLS, which was the first verified standard-compliant implementation. However, it still involved hand-written proofs and was more of an experimental platform. Enter Everest: They want to tick all the boxes by providing verified drop-in replacements for the HTTPS ecosystem. It covers the whole range from security definitions to code with good performance and side-channel protection.

As an example, Cédric presented Poly1305, a MAC that uses arithmetic modulo $2^{130}-5$ and forms part of the upcoming TLS 1.3 specification. Unsurprisingly, there have already been found bugs in OpenSSL's implementation. Project Everest have implemented Poly1305 with ~3,000 lines of code in Low*, a subset of F* (a functional language) that allows both C-style programming (with pointer arithmetic) as well as verification. Compiling this code with KreMLin (another output of Project Everest) results in machine code that is as fast as hand-written C implementations. The same holds for ChaCha2 and Curve25519.

However, hand-written assembly is still faster. The project aims to catch up on this with Vale, which was published at Usenix this year. Vale promises extensible, automated assembly verification and side-channel protection.

So what is the way forward? TLS 1.3 is on the horizon, bringing various improvements at the cost of a considerable re-design. This requires new implementations, and the project hopes to be first to market with an efficient and verifiable one.

For the rest of talk, Cédric gave more details on how F* naturally lends itself to security games, and how they proved concrete security bounds for the TLS 1.2 and 1.3 record ciphersuites.

All in all, I think it was a great example for an invited talk at Crypto, putting cryptography in a bigger context and bringing work that isn't necessarily on a cryptographer's radar to our attention.

Crypto 2017 - LPN Decoded

Crypto 2017 kicked off this morning in Santa Barbara. After a quick eclipse-watching break, the lattice track ended with the presentation of LPN Decoded by Andre Esser, Robert Kübler, and Alexander May.

Learning Parity with Noise (LPN) is used as the underlying hardness problem in many cryptographic protocols. It is equivalent to decoding random linear codes, and thus offers strong security guarantees. The authors of this paper propose a memory-efficient algorithm to the LPN problem. They also propose the first quantum algorithm for LPN.

Let's first recall the definition of the LPN problem. Let a secret $s \in \mathbb{F}_2^k$ and samples $(\mathbb{a}_i, b_i)$, where $b_i$ equals $\langle \mathbb{a}_i,s \rangle + e_i$ for some $e \in \{0,1\}$ with $Pr(e_i=1)= \tau < 1/2$. From the samples $(\mathbb{a}_i,b_i)$, recover $s$. We see that the two LPN parameters are $k$ and $\tau$. Notice that this can be seen as a sub-case of the Ring Learning with Errors problems; in fact, LWR originated as an extension of LPN.

If $\tau$ is zero, we can draw $k$ independent samples and solve for $s$ by Gaussian elimination. This can also be extended to an algorithm for $\tau < 1/2$, by computing a guess $s'$ and testing whether $s'=s$. This works well for small $\tau$, for example $\tau = 1/\sqrt{k}$, used in some public key encryption schemes. Call this approach Gauss.

For larger constant $\tau$, the best current algorithm is BKW. However, although BKW has the best running time, it cannot be implemented for even medium size LPN parameters because of its memory consumption. Further, BKW has a bad running time dependency on $\tau$. Both algorithms also require many LPN oracle calls. 

The authors take these two as a starting point. They describe a Pooled Gauss algorithm, which only requires a polynomial number of samples. From those, they look for error-free samples, similar to Gauss. The resulting algorithm has the same space and running time complexity, but requires significantly less oracle calls. It has the additional advantage of giving rise to a quantum version, where Grover search can be applied to save a square root faction in the running time. 

They then describe a second hybrid algorithm, where a dimension reduction step is added. Thus Well-Pooled Gauss proceeds in two steps. First, reduce the dimension $k$ to $k'$ (for example, using methods such as BKW). Then, decode the smaller instance via Gaussian elimination. The latter step is improved upon by using the MMT algorithm.

For full results, see the original paper. Their conclusion is that Decoding remains the best strategy for small $\tau$. It also has quantum optimisations and is memory-efficient. The Hybrid approach has in fact no advantage over this for small values of $\tau$. For larger values however, they manage to solve for the first time an LPN instance of what they call medium parameters - $k = 243$, $\tau = 1/8$ - in 15 days.

Eurocrypt 2017 - Parallel Implementations of Masking Schemes and the Bounded Moment Leakage Model

Side-channel analysis made its way into Eurocrypt this year thanks to two talks, the first of which given by François-Xavier Standaert on a new model to prove security of implementations in. When talking about provable security in the context of side-channel analysis, there is one prominent model that comes to mind: the d-probing model, where the adversary is allowed to probe d internal variables (somewhat related to d wires inside an actual implementation) and learn them. Another very famous model, introduced ten years later, is the noisy leakage model in which the adversary is allowed probes on all intermediate variables (or wires) but the learnt values are affected by errors due to noise. To complete the picture, it was proved that security in the probing model implies security in the noisy leakage one.

The work of Barthe, Dupressoir, Faust, Grégoire, Standaert and Strub is motivated precisely by the analysis of these two models in relation to how they specifically deal with parallel implementation of cryptosystems. On one hand, the probing model admits very simple and elegant description and proofs' techniques but it is inherently oriented towards serial implementations; on the other hand, the noisy leakage model naturally includes parallel implementations in its scope but, admitting the existence of noise in leakage functions, it lacks simplicity. The latter is particularly important when circuits are analysed with automated tools and formal methods, because these can rarely deal with errors.

The contribution of the paper can then be summarised in the definition of a new model trying to acquire the pros of both previous models: the Bounded Moment leakage Model (BMM). The authors show how it relates to the probing model and give constructions being secure in their model. In particular, they prove that BMM is strictly weaker than the probing model in that security in the latter implies security in the former but they give a counterexample that the opposite does not hold. The informal definition of the model given during the talk is the following:

An implementation is secure at order o in the BMM if all mixed statistical moments of order up to o of its leakage vectors are independent of any sensitive variable manipulated.

A parallel multiplication algorithm and a parallel refreshing algorithm are the examples brought to show practical cases where the reduction between models stated before holds, the statement of which is the following:

A parallel implementation is secure at order o in the BMM if its serialisation is secure at order o in the probing model.

The falsity of the converse is shown in a slightly different setting, namely the one of continuous leakage: the adversary does not just learn values carried by some wires by probing them, but such an operation can be repeated as many times as desired and the probes can be moved adaptively. Clearly this is a much stronger adversary in that accumulation of knowledge over multiple probing sessions is possible, which is used as a counterexample to show that security in the continuous BMM does not imply security in the continuous probing model. The refreshing scheme mentioned above can easily be broken in the latter after a number of iterations linear in the number of shares, but not in the former as adapting the position of the probes does not help: an adversary in the BMM can already ask for leakage on a bounded function of all the shares.

Both slides and paper are already available.

Eurocrypt 2017: On dual lattice attacks against small-secret LWE and parameter choices in HElib and SEAL

This morning, Martin gave a great talk on lattice attacks and parameter choices for Learning With Errors (LWE) with small and sparse secret. The work presents new attacks on LWE instances, yielding revised security estimates. This leads to a revised exponent of the dual lattice attack by a factor of 2L/(2L+1), for log q = Θ(L*log n). The paper exploits the fact that most lattice-based FHE schemes use short and sparse secret. We will write q to denote the LWE modulus throughout.

Let's first have a look at the set-up. Remember LWE consists of distinguishing between pairs (A, As+e) and (A,b). In the first instance, A is selected uniformly at random and b is selected from a special (usually Gaussian) distribution. In the second one, both A and b are uniformly random. Selecting s, as this work shows, is perhaps trickier than previously thought. Theory says that, in order to preserve security, selecting a short and sparse secret s means the dimension must be increased to n*log_2(q). Practice says just ignore that and pick a small secret anyway. More formally, HElib typically picks a secret s such that exactly h=64 entries are in {-1,1} and all the rest are 0. SEAL picks uniformly random secrets in {-1,0,1}.

We also recall that the dual lattice attack consists of finding a short vector w such that Aw = 0, then checking if

<Aw, (As+e)w> = <w,e>

is short. If we are in the presence of an LWE sample, e is short, so the inner product is short. Short*short = short, as any good cryptographer can tell you.

The improvements presented in this paper rely on three main observations. Firsly, a revised dual lattice attack is presented. This step is done by adapting BKW-style algorithms in order to increase efficiency and can be done in general, i.e. does not depend on either shortness or sparseness of the secret. It is achieved by applying BKZ to the target basis, then re-randomising the result and applying BKZ again, with different block size.

The second optimisation exploits the fact that we have small secrets. We observe that we can relax the condition on w somewhat. Indeed, if s is short, then finding w such that Aw is short instead of 0 is good enough. Therefore, we look for vectors (v,w) in the lattice

L = {(y,x): yA = x (mod q)}.

Now in small secret LWE instances, ||s||<||e|| and so we may allow ||v||>||w|| such that

||<w,s>|| ≈ ||<v,e>||.

Finally, the sparsity of the small secret is exploited. This essentially relies on the following observation: when s is very sparse, most of the columns of A become irrelevant, so we can just ignore them.

The final algorithm SILKE is the combination of the three above steps. The steps are the following.

• Perform BKZ twice with different block sizes to produce many short vectors
• Scale the normal form of the dual lattice
• If sparse, ignore the presumed zero columns, correct for mistakes by checking shifted distribution

As usual, Martin wins Best Slides Award for including kittens.

Is Your Banking App Secure?

Last week I was in Malta for Financial Cryptography and Data Security 2017 to present my recent work on securing the PKCS#11 cryptographic API.

One talk that stood out for me was by researchers from the University of Birmingham, who looked for vulnerabilities in the mobile apps provided by major UK banks.

Sadly, they found major weaknesses in apps from 5 of the 15 banks they investigated.

Several apps use certificate pinning, where the app hard-codes a certificate from a trusted CA and only accepts public keys that are signed by the pinned certificate.
This is good practice, as an attacker can add their own certificate to the phone's trust store, but it won't be accepted by the app.
However, two Android apps (for Natwest and Co-op) accepted any public key signed by the pinned certificate, without checking the domain name!
So the attack works as follows:

1. Purchase a certificate for a domain you own from the trusted CA
2. The app will accept your public key with this certificate
3. Man-in-the-middle all the encrypted traffic between the user and their bank.

Curiously, the authors note: "Co-op [...] hired two penetration testing companies to test their apps, both of which had missed this vulnerability". It seems odd that such an obvious mistake wasn't picked up in testing.

The group also found that several banks - Santander, First Trust and Allied Irish - served adverts to their app users over unencrypted HTTP, meaning an attacker could spoof these ads and mount a phishing scam, perhaps by displaying a fake 'security warning' and directing users to re-enter their account details on a malicious page. It was pointed out in the talk that we're much more likely to 'feel safe' within an app (and hence trust all the content we see) than, say, visiting a webpage using a laptop, so this kind of in-app phishing scam could be very effective.

There are even more exploits described in the paper.

It was refreshing to hear that the vulnerable banks responded well to the disclosures made by the Birmingham group and patched their apps as a result. But I'm a little baffled that these basic errors were ever made in such security critical applications.

PKC 2017: Kenny Paterson accepting bets on breaking TLS 1.3

The member of the TLS 1.3 working group is willing to bet for a beer that the 0-RTT handshake of TLS 1.3 will get broken in the first two years.

In his invited talk, Kenny managed to fill a whole hour on the history of SSL/TLS without even mentioning symmetric cryptography beyond keywords, thus staying within the topic of the conference. Despite all versions of SSL being broken to at least some degree, the standardised TLS became the most import security protocol on the Internet.

The core part of TLS is the handshake protocol, which establishes the choice of ciphers and the session key. Kenny highlighted the high complexity stemming from the many choices (e.g., using a dedicated key exchange protocol or not) and the possible interaction with other protocols in TLS. Together with further weaknesses of the specification, this created the space for the many attacks we have seen. On the upside, these attacks express an increased attention by academics, which comes together with an increased attention by the society as whole. Both have laid the ground for improvements in both the deployment and future versions of TLS. For example, the support of forward secrecy has increased from 12 percent to 86 according to SSL pulse.

Turning to concrete attacks, most important in the area of PKC is the Bleichenbacher attack published already at Crypto 1998 (a human born then would a considered a full adult at the conference venue now). Essentially, it exploits that RSA with the padding used in TLS is not CCA-secure, and it recovers the session key after roughly $2^{20}$ interactions with a server. Nevertheless, the TLS 1.0 specification published shortly after Bleichenbacher's publication incorporates the problematic padding (recommending mitigation measures), and later versions retain it for compatibility. The DROWN shows the danger of this by exploiting the fact that many servers still offer SSLv2 (about 8% of Alexa top 200k) and that it is common to use the same key for several protocol versions. An attacker can recover the session key of a TLS session by replaying a part of it in an SSLv2 session that uses the same key.

On a more positive note, Kenny presented the upcoming TLS 1.3, which is under development since 2014. It addresses a lot of the weaknesses of previous versions by involving academics from an early stage and doing away with a lot of the complexity (including reducing options and removing ciphers). It furthermore aims to decrease latency of the handshake by allowing the parties to send encrypted data as early as possible, reducing the round trip time to one in many cases. The goal of low latency has also led to the inclusion of QUIC, which provides zero round trip time, that is, the client can send data already in the first message when resuming a session. However, QUIC is not fully forward-secure and therefore confined to a separate API. Nevertheless, Kenny predicts that the sole availability will be too tempting for developers, hence the bet offered.

Concluding, he sees three major shifts in TLS this far: from RSA to elliptic-curve Diffie-Hellman, to Curve25519, and away from SHA-1 in certificates. A fourth shift might happen with the introduction of post-quantum algorithms such as Google's deployment of New Hope. Less optimistically, he expects that implementation vulnerabilities will continue to come up.

Update: An earlier version of this post mentioned the non-existing Curve255199 instead of Curve25519, and it attributed New Hope to Google.

Tools for proofs

Security proof  for even simple cryptographic systems are dangerous and ugly beasts. Luckily, they are only rarely seen: they are usually safely kept in the confines of ``future full-versions'' of papers, or only appear in cartoon-ish form, generically labelled as ... ``proof sketch". 

The following two quotes frame the problem in less metaphorical terms. 

``In our opinion, many proofs in cryptography have become essentially  unverifiable. Our field may be approaching a crisis of rigor".

                                              Bellare and Rogaway (cca 2004)

``Do we have a problem with cryptographic proofs? Yes, we

do [...] We generate more proofs than we carefully verify

(and as a consequence some of our published proofs are

incorrect)". 

       
                                                                     Halevi (cca 2005)


Solutions developed by cryptographers e.g. compositional reasoning and the game-hopping technique, help to structure proofs and reduce their complexity and therefore alleviate to some extent the pain of having to develop rigorous proofs. Yet, more often than not proofs are still sketchy and shady.

There is help that comes from the programming languages community which has a long experience with developing tools for proving that programs work correctly and...cryptographic systems are just programs. Recent progress,  e.g. automated verification of parts of TLS, fully verified security proofs of implementation masking schemes to defeat leakage, is impressive and exciting.  More work is under way. 

If you want to learn more about how can you get someone else to do the proofs for you or, more realistically, learn about what existent tools can currently do, what they could do in the future, and discuss what is needed and which way to go, then you should attend the 

Workshop on Models and Tools for Security Analysis and Proofs
 -- Paris, 29th of April; co-located with EuroSnP and Eurocrypt --

which the Bristol Crypto group helps organize. The workshop features as speakers some of the most prominent researchers that are contributing to this direction. You can register for the workshop HERE. Early registration ends March 31st!


But wait...there is more. If you want to explore this area beyond what a one-day workshop allows, then you should consider attending the

Summer School on Models and Tools for Proofs 

-- Nancy, France, July 10th - 13th --

See you all in Paris and/or Nancy!

Homomorphic Encryption API Software Library

The Homomorphic Encryption Application Programming Interface (HE-API) software library is an open source software library being developed as part of the Homomorphic Encryption Applications and Technology (HEAT) project, and is available here. The main purpose of this software library is to provide a common easy-to-use interface for various existing Somewhat Homomorphic Encryption (SHE) libraries. Limited support for fixed-point arithmetic is also provided by this library. Note that the HE-API library is still a work in progress.

Fully Homomorphic Encryption (FHE) is a cryptographic primitive that allows meaningful manipulation of ciphertexts. In spite of several recent advances, FHE remains out of practical reach. Hence a reasonable restriction to make is to limit the set of evaluated circuits to a specified subclass, usually determined by the multiplicative depth of the circuit. Such encryption schemes are called as SHE schemes.  Various libraries such as HElib, SEAL, FV-NFLlib, HElib-MP, etc., are already available that implement these SHE schemes.

The purpose of this HE-API software library is to provide a common, generic, easy-to-use interface for various existing libraries that implement SHE schemes. The SHE libraries that are currently integrated in the HE-API library are HElib and FV-NFLlib. It may be noted that the FV-NFLlib library is itself an outcome of the HEAT project. At a high-level, the HE-API software library abstracts out the technicalities present in the underlying SHE libraries. For instance, the HElib library implements the BGV SHE scheme, while the FV-NFLlib implements the FV SHE scheme. Needless to say, the syntax for various classes and routines in the individual libraries will be different, though the underlying semantics are very similar. The HE-API library integrates the underlying SHE libraries under a single interface, thereby shielding the user from syntactic differences. Another feature of the HE-API library is that it contains minimal, yet complete, set of routines to perform homomorphic computations. The design of this library is motivated by the ease of use for non-experts.

Supported Data Types

The following application data types are supported by the HE-API software library. 

• Boolean
• Unsigned long integers
• GMP's arbitrary precision integers class: mpz_class
• Polynomials with coefficients of type: unsigned long integers or mpz_class
• Vectors of : unsigned long integers or mpz_class
• Fixed-point numbers

Note that all the data types and routines described above may not be currently supported by every underlying SHE library.

RWC 2017 - Secure MPC at Google

This talk was given by Ben Kreuter and its focus was on the apparent disparity between what we research in academia versus what is required in the real world, specifically in the field of multi-party computation (MPC). MPC is the idea of allowing multiple parties to compute some function on their combined input without any party revealing anything about their input to the other parties (other than what can be learnt from the output alone).

While significant work has been done on making MPC efficient in practice (for example, the work of Yehuda Lindell et al. on high-throughput MPC which was presented by Lindell in the preceding talk), the focus tends to be on generic protocols (e.g. general logic circuits) with strong security guarantees (e.g. malicious security), which invariably leads to large computational overhead. In practice, we usually require only specific protocols, which can therefore be optimised, and comparatively weak security guarantees.

In the real world, network cost is the salient factor, rather than the speed of the protocol, since the parties who are involved in a computation often have to use networks (such as the Internet) which are being used by many other people at the same time and cannot make the best use of the network's full capabilities. The MPC at Google is about computation amongst, for example, mobile phones, laptops and servers; this introduces issues like battery constraints and the possibility of the computation not completing; these considerations, firmly grounded in the real world, are important when developing MPC techniques in research.

Business applications

A large portion of Google's revenue is generated by advertising: the tech giant, well-known for its aptitude for accurately determining users' desired search results even when queries are expressed ineloquently, specialises in creating personalised adverts to its wide spectrum of users. The efficacy of an advert is generally measured by the proportion of viewers of it who later become customers. Clearly this can be done by businesses comparing their database of customers' transactions with Google's databases of who has been shown which adverts. This, however, would be an invasion of privacy: instead, Google and the business can do MPC: more specifically, a private set intersection protocol.

In a private set intersection protocol, the parties involved compute how large the intersection is amongst the sets input by each party, or even some function on those elements in the intersection. So if the business and Google compute a private set intersection protocol on their data, they can determine how well the advertising went.

Roughly speaking, the MPC Google does in the real world is as follows: Google has a set $\{g_1,g_2,...,g_n\}$ of field elements which encodes a set of people who have been shown an advert for a certain product, and a business has a set $\{b_1,b_2,...,b_m\}$ of field elements which encodes a set of people who have been sold the product in question; Google raises each of its elements to a power $G$ and sends the set $\{g_1^G,g_2^G,...,g_n^G\}$ to the business. The business does the same with its elements for some exponent $B$ to get $\{b_1^B,b_2^B,...,b_m^B\}$, encrypts a set of binary vectors under Paillier encryption (which is additively homomorphic), one corresponding to each element in its set, encoding some other property of the sales (like the amount paid), and also computes the set $\{g_1^{GB},g_2^{GB},...,g_n^{GB}\}$. The business sends Google the set of pairs $\{(b_1^B,P(v_1)),(b_2^B,P(v_2)),...,(b_m^B,P(v_m))\}$ along with $\{g_1^{GB},g_2^{GB},...,g_n^{GB}\}$, and Google computes $\{b_1^{GB},b_2^{GB},...,b_m^{GB}\}$ and adds together all encrypted vectors $P(v_i)$ for which there exists some $j$ such that $g_i^{GB}=b_j^{GB}$. It sends this ciphertext back to the business, which decrypts and interprets the result.

This protocol is very simple, and it is only passively secure (in which players are assumed to execute the protocol faithfully but will possibly try to learn things by inspecting their communication transcripts). An interesting, perhaps somewhat orthogonal concern, to how we approach research from an academic point of view is that it is important that we can convey the security and efficiency of our protocols to lawyers, managers and software engineers who will eventually be sanctioning, authorising or implementing the protocols. "The lawyers are interesting because you can show them a proof, and two plus two equals four is a negotiable statement here... managers usually trust your expertise...and software engineers are the worst because they already assume [the protocol] is impossible."

An alternative solution using garbled circuits was explored in the recent past, but it turned out that their use required some subtle assumptions regarding the computation and communication which would have made the protocol impractical.

Future work would involve getting a (not too much more expensive) maliciously secure protocol and developing the use of the homomorphic encryption to allow different functions to be computed on the data in the intersection.

Consumer applications

The Android keyboard app by Google, Gboard, logs what a user types so that it can guess words for auto-completing in the future. This data could be used for training machine learning models, and merging results from many local models would enable the formation of guessing algorithms that work well for everyone. However, to do this, the server would need to receive a set large dataset of words typed by a user from each phone so that this processing could be done. Clearly there is an issue of privacy here; moreover, there is also potentially a differential privacy issue.

This is clearly a good situation in which to use MPC. Each party masks their data using a basic additive secret-sharing scheme: if each party has a vector to input, for every coordinate, every pair of parties agrees on some random field element, one subtracts and one adds this to that coordinate of their vector. When the parties send this to Google, the masks will therefore cancel when added together.

In practice,they use a PRG and perform a key exchange (in which one key is given to each pair of parties, for every possible pair) at the beginning to achieve the same effect but with much smaller communication overhead. They also have a trick for dealing with device failures (which is important given the application).


This talk provided helpful and relevant insight into the the importance of matching what we research with what we require in the real world, which is, after all, one of the main reasons for having conferences such as Real World Crypto. Many of the talks are available to watch online here, and I would highly recommend doing so if interested.

RWC 2017 - Is Password Insecurity Inevitable?

Fresh back from an enlightening trip across the pond, I wanted to write about one of my favourite talks, all about password (in)security, from this year's Real World Cryptography conference.

As we know:

1. Passwords protect everything.
2. Passwords are terrible.

But happily, Hugo Krawczyk from IBM Research spoke about some great new work to resolve these two seemingly incompatible statements. There were a lot of details in the talk that I'll have to miss out here (slides are available online). In particular, I'm going to focus on 'Part I: Take the burden of choosing and memorising passwords off humans'.

The basic idea - this isn't new - is to have the user memorise a single master password that they use to access a password store. Then the password store derives unique pseudorandom passwords for each service the user wants to access (Google, Facebook, etc.) The problem with this solution is that the password store becomes a single point of failure: if it is compromised, then an offline dictionary attack to find the master password will compromise all of the user's accounts at once.

Krawczyk et al. suggest an improvement: SPHINX, which amusingly stands for "a password Store that Perfectly Hides from Itself (No eXaggeration)". The first idea is for the password store to not keep hold of (even a hash of) the master password - instead it has an independent secret key $k$, and any time the user wants to log in to a service $S$, they send the master password $pwd$ to the store, the store computes a PRF $PRF(k, pwd | S)$ and this will be sent to $S$ as the user's password for $S$. This means that if the store is compromised, the master password and the account passwords can't be learned unless the user communicates with the store. So this works well if the store is in local, offline hardware, where the user is unlikely to use the store after it is compromised by an attacker.

However, the authors go further and replace the PRF with an oblivious PRF. This means the store computes an "encrypted" version of $PRF(k, pwd | S)$ from an "encrypted" $pwd|S$, so doesn't learn the plaintext values of the master password or the service password. In practice this can be achieved by the user (i.e. the user's machine) hashing the string $pwd | S$ into an element $g$ of a Diffie-Hellman group, then computing $h = g^r$, where $r$ is a fresh, random exponent, and sending $h$ to the password store. The store's secret key is an exponent $a$, so it computes $h^a$ and sends this back to the user. The user removes the blinding exponent $r$ (i.e. computes $(h^a)^{r^{-1}} = g^a$) and the result is the unique password for $S$. Now even when the password store is compromised and even if the user communicates with the store, the master password and the account passwords can't be learned.

In principle an attacker could recover all the account passwords by compromising both the password store and a service $S$, learning the secret key $a$ and the service password $g^a$, computing $g = H(pwd|S)$ and perfoming an offline dictionary attack to find $pwd|S$. Then for any other service $S'$, the password can be computed via $H(pwd|S')^a$. But as long as $S$ follows good practice and only stores a hash $H'(g^a)$ of the service password, this attack fails: an offline dictionary attack to recover $g^a$ is unfeasible as it's essentially a random group element.

There are no particularly expensive computations involved in using SPHINX, the communication between the user and SPHINX does not need to be secure (so it could be online somewhere) and the store will work regardless of what password protocol is used by the service, so it's extremely flexible. SPHINX therefore strikes me as both useful and practical, which is surely the definition of Real World Cryptography.

RWC 2017 - Post-quantum cryptography in the real-world

A new year takes off and, along with it, thousands of resolutions are formulated. Although I am not the right person to talk about them (my diet will begin next Monday), I wish to discuss a resolution that the cryptographic community as a whole has set for itself in this 2017. Because that's what people do at Real World Crypto (RWC): they talk about new threads, topics could be worth exploring during the new year, directions for their researches and interests. This year, for the first time in RWC, post-quantum cryptography (PQC) was given an entire session, clear sign that time is changing and the moment has come to bring the discussion to the real world. The message is clear: even if quantum computers are not popping up in tomorrow's newspapers, we can't postpone any longer.

A very simple reason for this was given by Rene Peralta, of the NIST PQC team, during the overture of the session: standardisation takes time, up to seven years if we start right now, and full transition takes even longer. I found Rene's presentation to be neat and direct: our public-key cryptography fails against quantum computers and our symmetric one needs some (non-drastic) modifications. The resolution is to "start thinking about it this year, possibly by November 30^th, 2017". However, a question arises quite naturally: are we ready?

The other three talks of the session tried to answer in the affirmative. Among the several PQC proposals that are around in theoretical papers, two made their ways into RWC: the well-stablished lattice-based cryptography and the new-born isogeny-based cryptography, which nevertheless carries the pride and sympathy of ECC.

Lattices and funny names: NewHope and Frodo and Crystals

Lattice-based cryptography has three representatives in the run for PQC schemes. Valeria Nikolaenko showed two: the first one is called NewHope and is a key agreement protocol based on the hardness of Ring-LWE. The latter is a problem very favourable to applications because it combines sound theoretical security (worst-case to average-case reduction) to fast implementations thanks to specific choices of parameters which allow for speed-ups in the computations: NewHope turns out to be even faster than ECC and RSA, but at the price of a larger communication. However, there are some concerns on the security of LWE when the ring structured is added. Thus, Frodo ("take off the ring") is designed to achieve the same goal using only standard LWE. The drawback is a degradation in performance, since the tricks hinted above cannot be used anymore and keys are generally bigger.

The third lattice-based scheme was presented by Tancrede Lepoint and is a suite called Crystals. This is based on yet another kind of lattices: module lattices, for which it is also known a worst-case to average-case reduction. These are less structured lattices (hence possibly calming down the detractors of ring structure) in which similar implementation speed-ups are possible: the timing is indeed comparable to NewHope's, while the communication is improved.

"Make elliptic curves great again"

Michael Naehrig presented a new proposal for PQC: do you remember curves with plenty of small subgroups where to easily solve the discrete logarithm problem? Now they come in handy again: all the subgroups (of order 2 and 3) are considered to be nodes of a graph, whose edges are the isogenies (a.k.a. bijetive homorphisms between curves). In this new context, given two curves in the graph, it is difficult to come up with the isogeny linking the two. However, such a new approach doesn't really stand against other solutions: keys are small but performance is not a pro (so to speak).

RWC 2017 - Erasing Secrets from RAM

One of my favourite talks from the Real World Crypto 2017 conference was given by Laurent Simon, on Erasing Secrets from RAM.

In short, it was found that in practice, many non-malicious programs handling keys and other sensitive data do not erase the RAM correctly. This would allow an attacker (that has access to all of a system's volatile memory and CPU state) access to any unerased sensitive data.

It was thought that compiler optimisation played a part in the lack of erasion. Take the code below:

void sensitive_function(...) {

   u8 sensitive_buffer[KEY_MAX] = "\0";
   ...
   zeromem(sensitive_buffer, KEY_MAX);

}

The compiler may choose to remove the zeromem line, as the sensitive_buffer is going out of scope anyway. This would leave sensitive data on the RAM, unbeknownst to the programmer.

So, the paper presents a tool that allows developers to mark sensitive variables in their code, and then see (post-compilation) any potential leakage of sensitive data.

They call this tool Secretgrind, based off the popular Valgrind.

Anyway, as it turns out, the compiler optimisation problem mentioned above wasn't actually a problem in practice - they didn't once encounter this problem in all their testing. Instead, the majority of sensitive leaks were down to developers' mistakes; they had forgotten to erase sensitive variables on both the stack and the heap.

There were a few issues with IO API's caching optimisations, though - such as when you read a line from a PEM file using mmap, it often loads the whole file into memory to save you the time. However, this is not immediately obvious, and when you go to delete the line from RAM, the rest of the file is still in memory!

Laurent concluded the talks saying Secretgrind was still in development, and although referring to it as a 'hack' (due to it's fragility), wishes for it to be used to "help you guys check your code".

RWC 2017 - A Formal Security Analysis of the Signal Messaging Protocol

Real World Crypto 2017 kicked off yesterday in New York City. This afternoon, Luke Garratt presented his and his colleagues' work, A Formal Security Analysis of the Signal Messaging Protocol. The signal protocol is used by the Facebook messenger app, WhatsApp and the Signal app (to name but a few). It is therefore surprising that there had been no formal security analysis, and their work addresses this issue. The paper is motivated by the questions

What should Signal achieve?

and

Does it?

Or, put in more modern language (spoiler alert),

Why what does is 

Let's first look at what kind of security we can hope for. We speak of Forward Secrecy, which ensures that past messages are not compromised even if the communication is at some point in the future. This is to prevent an adversary storing all the conversations and waiting until an attack is successful to then recover all the communication. 

Post-compromise security pushes this even further. If we have post-compromise security, not only past conversations are not compromised, but also future ones. The Signal protocol achieves this using a technique called ratcheting, which involves session keys being updated with each message sent. Why is this useful? Well, it makes the adversary's life much harder. In order to attack a system with post-compromise security, an adversary must obtain long-term keys and immediately attack and continue attacking if they want to compromise future sessions. As opposed to forward security, where an adversary would obtain a long-term key and wait for an interesting target to launch a man-in-the-middle attack (e.g. TLS-DHE) or to no forward security, where they would just store all ciphertext traffic until they obtain a long-term key and decrypt everything (e.g. TLS-RSA). Times are hard. 

Their security model captures: 

• Full network control by the adversary;
• Perfect forward secrecy;
• Post-compromise security;
• Key compromise impersonation attacks;
• Some (but not all) random numbers compromise.

Their proof is too long to be featured in this blog post, but Luke promises it is tedious rather than complex. Their conclusion? So far, so good. 

In terms of limitations, they note that they have not analysed any implementation of the protocol, so this is a theoretical analysis only. They also assume an honest key distribution and have not (yet) considered multiple devices. 

Study Group - All Your Queries Are Belong to Us: The Power of File-Injection Attacks on Searchable Encryption

Published at USENIX 2016 and written by Yupeng Zhang, Jonathan Katz, and Charalampos Papamanthou from University of Maryland this paper proves again how difficult is to get proper security in searchable encryption.

Some of you may wonder why I chose this paper. Reason is that I wanted to get a grasp of how research looks like outside of the MPC area at a top security conference. And what other conference to pick than this year's USENIX?

After I finished reading this paper I realised a cool thing: that you need little to none a priori knowledge about Searchable Encryption (SE) [1]. Fortunately I didn't find myself diving into the literature of SE in order to fully understand the ideas presented there - how many crypto papers have you read and still be able to say this once you're done with them?

Let's introduce first the notion of SE. The setup consists in 2 parties, a client and an encrypted mail server. The client wants to obtain all e-mails which have 'alice' and 'work' as tags. He then computes a token in a deterministic manner, then the server replies with the e-mails  corresponding to the query:

To ease the presentation we can assume that instead of e-mails the server will respond with identifiers (urls).

An adversary wins if he breaks the token privacy, namely it recovers keywords from tokens.

A file injection attack is when an attacker has encrypted e-mails of his choice on the server and can see the tokens which are queried by the client. More simple, the server behaves in a honest-but-curious way and stores encrypted e-mails of his choice by spamming the client.

From a first sight this seems harmless. But if you combine it with some leaked e-mails (lots of them these days) the authors managed to have a 70% recovery rate for a keyword from a token with only 20% of leaked files. This was tested using an Enron email dataset of ~30000 emails. [2]

To understand how this works it's important to look at a simpler attack:

Binary search attack.

Suppose an adversary has a keyword universe $K$, with size $|K|$ a power for $2$. He can inject $\log_2{|K|}$ files and then recover keywords from every token with $100\%$ success.

Each file $F_i$, $i \in [0, \log_2{|K|}] $ contains keyword $k_j$ for which $j$ has $i$'th most significative bit equal to $1$. To see why this works let's look at an example inspired from the paper.

In this case we have $|K| = 8$. Black cells are the keywords inserted in file $i$. Since the search result is $F_1$, $F_3$ ($101$) we conclude that the token was derived from the keyword $k_5$.
A server which inserts one email per day can execute an attack in $2$ weeks time for $10,000$ keyword universe.

Countermeasure.

One crucial observation is that we insert $|K|/2$ keywords per file. So one can think that limiting the keywords per file to some threshold $T << |K|/2$ will fix the problem.

The threshold countermeasure turns out to be ineffective. The authors proved this with an attack which works almost like the first one after splitting the keywords in chunks of size $2T$. (denoted as hierarchical search attack).

For $|K| = 5,000$ and $T=200$ an adversary should can inject $131$ files in order to recover keywords for every token.

Attacks with partial leakage.

Sometimes e-mails leak. And when that happens...SE is almost useless against adaptive injection attacks as the authors prove. We say adaptive because it needs a forward insecure SE scheme.

What if we have want to recover keywords from multiple tokens?

The idea is to compute the distribution for each keyword $f^*(k)$ from the leaked files. Then the assumption that $f^*(k)$ is close to the queried tokens distribution $f(t)$ is made - which will turn true latter.
Next, a small candidate set of keywords is chosen (also called ground truth) and files are injected using the binary search attack. The rest of the tokens are recovered by taking joint distributions with previous tokens.

For more attacks and countermeasures which fail I strongly recommend reading the article [3].

At the end of the article one can wonder if building secure SE schemes would really help against these access pattern attacks...The authors suggest that an interesting direction is to look into lower bounds on these attacks but this seems a far from trivial task.

[1] http://crypto.stanford.edu/~dabo/pubs/papers/encsearch.pdf
[2] https://www.cs.cmu.edu/~./enron/
[3] https://eprint.iacr.org/2016/172

What is SPDZ? Part 3: SPDZ specifics

This blog post is the final in a series of three in which we describe what SPDZ is. The first part can be found here, and the second here.

In this part, we consider what it is that makes SPDZ SDPZ.

SPDZ MPC

Using the term SPDZ is perhaps a little misleading. There are actually a few protocols which we call SPDZ because they are really just improvements on the original. As the online phase of SPDZ is already ‘pretty good’, optimisations to the protocol normally involve improving the offline phase.

In what follows, we talk about the methods used in SPDZ for preprocessing (generating the triples) and then explain some improvements that have been made.

Original SPDZ

Here we will outline some of the methods in the original SPDZ protocol, as in [5].

Preprocessing

In the first SPDZ paper, the authors show how to use somewhat homomorphic encryption (SHE) to generate the triples. This is a (pretty expensive) public key cryptosystem in which the encryption operation is ring homomorphic - that is, \[\textsf{Enc}_{\textsf{pk}}(a)\boxplus\textsf{Enc}_{\textsf{pk}}(b)=\textsf{Enc}_{\textsf{pk}}(a+b)\] and \[\textsf{Enc}_{\textsf{pk}}(a) \boxdot \textsf{Enc}_{\textsf{pk}}(b) = \textsf{Enc}_{\textsf{pk}}(a \cdot b)\]

(For those who know about FHE: we can do MPC with FHE, as you’d hope; however, at the time of writing no FHE scheme which is competitive with current secret-sharing MPC currently exists. SPDZ is what the authors call the ‘ideal use of FHE in MPC’: raw material is computed locally so that the online computations are minimal and communication overhead is small.)

During the offline phase, we also generate lots of random values which are used by the parties to provide input to the circuit.

Sacrificing

In order to check that the adversary has not introduced errors into the triples, for each triple we want to use we have to ‘sacrifice’ a triple by taking another triple from the preprocessing and checking that the third element is the product of the first two using Beaver’s method outlined in the previous post. Fortunately the method by which triples are generated is very efficient.

ZKPoPKs

Zero-Knowledge Proofs of Plaintext Knowledge (ZKPoPKs) are a way of ensuring parties encrypt as they should: when encrypting in the SHE scheme, there are conditions on the plaintext and noise that need to be met to ensure the ciphertext is well-formed.

MACs

Before any preprocessing computation takes place, the parties agree on some MAC key $\alpha$ with which they MAC all their data and which they secret-share amongst themselves so that no individual party knows the MAC key.

This MAC key is used to MAC all the data in the circuit. For each secret-shared value, there is also a secret-shared MAC on that value.

After the circuit has been evaluated, the parties open the secret corresponding to the circuit output and also the MAC on it, and check that the MAC is correct. If an actively corrupt party cheats at any point in the circuit evaluation, the final MAC check reveals this has happened. Note that this means the parties could be computing on invalid data.

In the original paper, each party also MACked the global MAC key with a personal MAC key. This was not needed in subsequent works.

Updated SPDZ: SDPZ2

While the online phase of SPDZ was (almost) the same, in SPDZ2 [4] the authors made significant changes to the offline phase of the original protocol. We outline the major changes here.

They solved an open problem left by the first paper in giving a secure protocol which generated a shared SHE decryption key.

They also offered various performance enhancements, including the introduction of ‘squares’ and ‘bits’ in the preprocessing, which allowed faster online computations of specific operations in the circuit.

A major improvement was allowing MACs to be checked on data before the end of the computation. This involved checking the MAC of a public value that depended on the underlying data. The big advantage of this is that the preprocessed data MACked under the same key can still be used after this MAC check, but not in the original protocol since the check reveals the key.

The new protocol had covert and malicious variants, the former having lesser parameter requirements. A covertly secure protocol ensures the adversary wins with probability at most $1/n$ for some $n \in \mathbb{N}$. The reason for this was that it was believed more closely to match the real-world situation.

New SPDZ: MASCOT

The most recent improvements to the SPDZ protocol is a work known as MASCOT [3]. As with SPDZ2, the online phase of SPDZ was left as it was; this paper improves on the offline phase, showing how to use oblivious transfer (OT) to achieve much faster preprocessed data than by using SHE.  It is authored by Bristol’s own Marcel, Emmanuela and Peter, and appeared at CCS last week.

(Oblivious transfer is a process in which one party inputs two strings and a second party chooses exactly one; this is done in such a way that the first party doesn’t know which string the second chose, and the second does not learn anything about the string it didn’t choose.)

The authors of MASCOT note that the heavy machinery of public-key cryptography required in the original SPDZ paper to generate the preprocessing (namely, the use of SHE) prevents the communication from being a bottleneck since the local computation takes so long.

The protocol is actively secure and performs significantly faster than previous SPDZ2 implementations, with much lower communication overhead.

The Future of SPDZ

MASCOT has only just been released, and there are already whispers of a new paper. Watch this space...

References

[1] B. Applebaum, Y. Ishai, and E. Kushilevitz. How to garble arithmetic circuits. 52nd FOCS, pp120–129. IEEE Computer Society Press, 2011

[2] D. Beaver. Efficient Multiparty Protocols using Circuit Randomisation. In J. Feigenbaum, editor, CRYPTO, volume 576 of Lecture Notes in Computer Science, pp420-432, Springer, 2012.

[3] R. Bendlin, I. Damgard, C. Orlandi, and S. Zakarias. Semi-homomorphic encryption and multiparty computation. In EUROCRYPT, pp169-188, 2011.

[4] I. Damgard, M. Keller, E. Larraia, V. Pastro, P. Scholl, N. P. Smart. Practical covertly secure MPC for dishonest majority - or: Breaking the SPDZ limits. In ESORICS (2013), J. Crampton, S. Jajodia, and K. Mayes, Eds., vol. 8134 of Lecture Notes in Computer Science, Springer, pp. 1–18.

[5] I. Damgard, V. Pastro, N. P. Smart, and S. Zakarias. Multiparty computation from somewhat homomorphic encryption. In Advances in Cryptology – CRYPTO 2012, volume 7417 of LNCS, pp643–662. Springer, 2012.

[6] I. Damgard and S. Zakarias. Constant-overhead secure computation of

boolean circuits using preprocessing. In TCC, pp621-641, 2013.

[7] M. Keller and E. Orsini and P. Scholl. MASCOT: Faster Malicious Arithmetic Secure Computation with Oblivious Transfer. Cryptology ePrint Archive, Report 2016/505, 2016.

[8] J. Buus Nielsen, P. Nordholt, C. Orlandi, and S. Burra. A new approach to practical active-secure two-party computation. In Reihaneh Safavi-Naini and Ran Canetti, editors, Advances in Cryptology CRYPTO 2012, volume 7417 of Lecture Notes in Computer Science, pp681-700. Springer Berlin Heidelberg, 2012.

[9] A. Shamir. How to Share a Secret. In Communications of the ACM, Volume 22 Issue 11, Nov. 1979, pp612-613.

[10] A. Yao. How to generate and exchange secrets. In SFCS '86 Proceedings of the 27th Annual Symposium on Foundations of Computer Science, pp162–167. IEEE, 1986.

Study Group - Dedup est Machina

The title of the paper is 'Dedup Est Machina: Memory Deduplication as an Advanced Exploitation Vector'.

I saw the paper being presented at the IEEE Security and Privacy Symposium, and it was one of the more enjoyable talks due to the nature of the exploit, and the enthusiasm of Erik Bosman presenting it. It also won a 'PWNIE' award for the 'most innovative research on hacking' at Black Hat USA this year.

The idea of the paper is simple - to combine the Memory Deduplication exploit with a Rowhammer attack.

They use this to craft a JavaScript-based attack against the Microsoft Edge browser, to gain arbitrary memory read and write access in the browser.

Memory Deduplication

The idea is to reduce the total memory footprint of a running system by sharing memory pages with identical content across independent processes.

A good example of this is the page cache in modern operating systems. The page cache stores a single cached copy of file system contents in memory, and shares the copy across different processes.

In running processes, two or more pages with the same content in run-time memory are always deduplicated, even if the pages are completed unrelated, and their equivalence is fortuitous.

So, in order to keep a single copy of a number of identical pages, the memory deduplication system needs to perform three tasks:

1. Detect. The system needs to detect memory pages with the same content, usually done at regular and predetermined intervals during normal system operations
2. Only Keep One. We only want to keep one physical copy, and return the others to the memory allocator; to do this, the deduplication system updates the page-table entries of the owning processes, so that the virtual addresses now point to a single shared copy; these entries are also marked as read-only to support copy-on-write semantics
3. Create Private Copy on Write. Specifically, when one of the owning processes writes to the read only page, a copy-on-write page fault occurs; at this point, the memory deduplication system can create a private copy of the page, and map it into the corresponding page table entries of the faulting process

On Windows 8.1 onwards, it's known as memory combining.

Memory Deduplication as a Side Channel

So you may have spotted how you could exploit this as a side channel. Writing to a shared page from any of the owning processes results in a page fault and a subsequent page copy.

Due to these additional expensive operations, a write to a shared page takes significantly longer compared to a write to a regular page (by significantly longer, we're talking up to one order of magnitude). This timing difference provides an attacker with a side channel to detect whether a given page exists in the system.

To do this, the attacker has four easy steps:

1. Craft a page with the exact same content as the target page
2. Wait for the memory deduplication system to do its thing
3. Write to the crafted page
4. Time how long it takes for the write operation to finish

If the write takes longer than a write to a non-deduplicated page, the attacker can conclude that a page with the same content exists.

So just using this, an attacker could detect whether a user was visiting a particular web page, or running a particular program.

Of course, false positives are possible due to noise etc, so the attacker can use redundancy (repeated attempts) to disclose the intended information in a reliable way.

At the moment, the memory deduplication side channel is looking a little bit pathetic, as it's a slow single-bit side channel that can only be used for leaking a limited number of bits from a victim process.

However, we can abuse memory deduplication to get some stronger primitives.

Primitives

Primitive #1: Alignment Probing

First up, we have a primitive that allows a byte-by-byte probing of secret information by controlling its alignment in memory. So, this primitive is applicable when the secret data has weak alignment properties.

Imagine in this case, the secret is a password stored in memory immediately after a blob of attacker-provided input bytes. What the attacker can do is provide some more input bytes, which effectively push the second part of the secret out of the target page. Now the attacker can brute force the first part of the secret (e.g. one byte) using deduplication.

All the attacker needs to do now is provide fewer input bytes to bring another portion of the secret back into the page. Just like before, the attacker uses deduplication to get the rest of the secret. For larger secrets, you just keep repeating these steps until all of the secret is known.

This alignment probing primitive is apparently very effective, and is used to disclose code pointers in Microsoft Edge, and a password hash in nginx.

Primitive #2: Partial Reuse

When the secret has strong memory alignment properties, the first primitive falls through.

Up next is a primitive that can be used when attacker-controlled input can partially overwrite stale secret data with predictable reuse properties. User-space memory allocators encourage memory reuse and do not normally zero out deallocated buffers for performance reasons.

This means that a target application often reuses the same memory page and selectively overwrites the content of a reused page with new data.

So, same set up as before. This time, when the attacker gives a larger input, it overwrites the first part of the secret. Now the attacker can brute force the second part using memory deduplication.

Once this is known, the attacker can brute force the first part of the secret, by deduplicating against a page without an overwritten secret.

This primitive is apparently fairly common in practice, and is used in the paper to break heap ASLR (Address Space Layout Randomisation) in nginx.

Primitive #3: Birthday Heap Spray

Last one. This primitive can leak a secret even when the attacker has no control over memory alignment or reuse. It relies on a generative approach that revolves around the birthday paradox, which we all know and love.

This primitive is applicable when the attacker can force the application to controllably spray target secrets over memory.
Up until now we have only assumed there is one secret we want to leak.

If a partially masked secret has $P$ possible values, we use memory deduplication to perform $1 \times P$ comparisons between the $P$ probe pages and the single target page - essentially brute forcing the secret.

For large $P$, a very large amount of memory is required, for the attack, and for the redundancy. However, if the attacker can cause the target application to generate many secrets, memory deduplication provides a much stronger primitive than brute forcing.

For instance, an attacker can generate a large number of secret heap pointers by creating a large number of objects from JavaScript, each referencing another object.

So, assume the attacker causes the application to generate $S$ such pages, each with a different secret pointer. The attacker now also creates $P$ probe pages, with $P$ being roughly the same size as $S$.

Each probe page uses the same encoding as the secret pages, except that, not knowing the secret pointers, the attacker needs to 'guess' their values. Each probe page contains a different guessed value. The idea is to find at least one of the probe pages matching any of the secret pages. So we have a classic Birthday Problem, where the secret and probe values are the birthdays.

Since memory deduplication compares any page with any other page in each deduplication pass, it automatically tests all $P$ probe pages against the $S$ target secret pages.

A hit on any of the $P$ possible values immediately exposes a target secret. This birthday primitive reduces the memory requirements of the attack by a factor of $S$. It's especially useful when leaking the location of randomised pointers. For exploitation purposes, it's not important which pointer the attacker hits, as long as at least one of them is hit.

This primitive is used to leak a randomised heap pointer in Microsoft Edge, and also used to craft a reliable Rowhammer exploit.

Rowhammer Exploit

Anyway, now we get to the good bit; the Rowhammer exploit. The Rowhammer bug was first publicly disclosed in June 2014 by Kim et al. However, it was not until March 2015 that someone published a working Linux kernel privilege escalation exploit using Rowhammer.

Essentially, Rowhammer is DRAM vulnerability that allows an attacker to flip bits in a (victim) memory page by repeatedly reading from other (aggressor) memory pages. These bit flips are deterministic, meaning that once a vulnerable memory location is identified, it is possible to reproduce the bit flip patterns by reading again the same set of aggressor pages.

There are two variations of the Rowhammer attack. Single-sided Rowhammer repeatedly activates a single row to corrupt its neighbouring rows' cells. Double-sided Rowhammer targets a single row by repeatedly activating both its neighbours rows. Although double-sided is more effective than single-sided in practice, it requires you to know not only the location of what you want to target, but also the two adjacent memory pages. Therefore, the authors of this paper choose to use the single-sided Rowhammer attack.

So, if you want to use the Rowhammer attack to do something useful, rather than to just corrupt random memory pages, then first you need to find a vulnerable memory location. To do this, you can allocate a large array filled with doubles. If you do it right, you can end up with an encoded double value such that all bits are set to 1.

Then, you find some eviction sets so you can bypass the cache, and you hammer 32 pages at a time (though it doesn't say where they get the number 32 from). By hammer, I mean you read from each page two million times before moving to the next page. Once you've gone through all 32 pages, you check the entire Array for bit flips. After scanning a sufficient number of pages, you will know which bits can be flipped at which offsets.

Next, you want to determine what to place in these vulnerable memory locations to craft the exploit. The idea behind the Rowhammer attack in this paper, is to place some data in the array which, after a bit flip, can yield a reference to a controlled counterfeit object.

So once we find out which bit is flipped, we can store a valid object reference at the vulnerable memory location and pivot to a counterfeit object with Rowhammer.

So that's the attack - using memory deduplication to leak the code and heap pointers, and Rowhammer to pivot them to a counterfeit object.  In practice, the paper says it takes about 30 minutes for the two deduplication passes and 508MB of memory. It also takes 1GB of memory to find bit flips, and 32MB to find the cache eviction sets. The timing of the Rowhammer depends on the vulnerable DRAM chips, so they say it varies from seconds to hours in practice.


Mitigation

So finally, the last part of the paper talks about how one would prevent attacks like this. A fairly obvious observation would be to just no deduplicate, as that prevents all exploits like this. However, deduplication contributes heavily to using physical memory efficiently, so we would like to keep it if possible.

This paper suggests that you only deduplicate zero pages. A zero page is a page with a starting address of zero, or rather, the series of memory addresses at the absolute beginning of a computer's address space. The paper doesn't go into much detail about why, but claims they've done experiments to show that it's nearly as efficient as full deduplication, but entirely eliminates the possibility for any attacks targeting memory deduplication.

Fin

So that's the end of paper, they finish by saying memory deduplication is dangerous in the wrong hands, and they've been working with their contacts at Microsoft to find an immediate solution.

I had a look round on the internet but no-ones has published anything further since the paper, so who knows?

Study Group: Towards Sound Fresh Re-keying with Hard (Physical) Learning Problems

The first study group on side-channel analysis in the academic year 2016/2017 is "Towards Sound Fresh Re-keying with Hard (Physical) Learning Problems" from Stefan Dziembowski, Sebastian Faust, Gottfried Herold, Anthony Journault, Daniel Masny and Francois-Xavier Standaert, and recently published at Crypto this year. There are mainly two reasons why I chose to present it: it is somehow close to my personal research and, generally speaking, I like to read of different areas influencing each other. In this particular case, learning problems which are usually associated to post-quantum cryptography (more specifically to lattice-based cryptography) have been used to prove a security result in leakage-resilient cryptography.

In one sentence and very loosely speaking, the content of the paper which has been covered in the study group can be summarised as: the authors construct a fresh re-keying scheme and show it is secure under the LPN assumption. The remaining of this blog post is structured as a number of questions and answers on the previous statement that will hopefully clarify and detail its meaning.

Q: what is a fresh re-keying scheme?

A: intuitively, a fresh re-keying scheme is a set of cryptographic algorithms which allows the generation of session keys from a master secret key and some publicly known randomness. A session key is used to encrypt a message and then discarded. The following diagram nicely represents what is going on.

Protecting block ciphers (or other cryptographic primitives) against DPA is an expensive procedure that introduces overhead in the design and whose security is often based on heuristic arguments. Such attacks exploit the dependence of multiple power traces with the same secret key to retrieve it. Hence, fresh re-keying schemes represent a solution that try to structurally avoid the threat by using each session key only once. Both the receiver and the sender can compute the same session key by means of the shared master secret key and some randomness associated with the ciphertext. At that point, the underlying block cipher should retain security only against SPA, while the algorithm the session key is computed by should resist DPA to prevent the adversary to gain information about the master secret key thanks to leakage. The overall scheme is designed to be more efficient than a one which applies DPA countermeasures directly to the block cipher because the GenSK algorithm is built to be easier to protect.

Q: what is the LPN assumption?

A: the Learning Parity with Noise (LPN) problem asks to find a vector $\mathbf{k}\in\mathbb{Z}_2^n$ given query access to an oracle that outputs $$ (\mathbf{r},<\mathbf{r},\mathbf{k}>\oplus e) \in \mathbb{Z}_2^n\times\mathbb{Z}_2 $$ where $\mathbf{r}\in\mathbb{Z}_2^n$ is a uniformly random vector and $e\leftarrow \mathcal{B}_\tau$ is an error bit drawn from the Bernoulli distribution of a fixed parameter $0<\tau<1/2$, i.e. $\mathbb{P}(e=1)=\tau$. The decisional version, instead, asks to distinguish the above distribution from the uniform one on $\mathbb{Z}_2^n\times\mathbb{Z}_2$. The first step in the overall proof is to define a similar version of such a problem, what the author call the Learning Parity with Leakage (LPL) problem, in which everything is the same bar the distribution of the noise, which is now taken to be the Gaussian distribution over the reals. Note that the second component of the LPL distribution is then a real number. It is shown that: $$ \text{LPN is hard} \Rightarrow \text{LPL is hard}. $$

Q: what does it mean for a fresh re-keying scheme to be secure?

A: since we deal with an environment where leakage exists, we should specify both what kind of security model the proof follows and what kind of leakage the adversary is allowed to query.

Security can be stated as the indistinguishability game in the above picture. In the real part on the left, the adversary plays with and queries a real oracle, that is to say an oracle that generates all keys and randomness according to the scheme to be secured. Hence also the leakage is computed on the sensitive variables. Instead, the right-hand side depicts the random game: an ideal oracle generates keys at random, which are then passed to a simulator which computes leakage on them. For the security claim to be valid, the (computational) adversary shouldn't distinguish the oracles she is playing with.
The leakage model is the $t$-noisy probing model. If you imagine a circuit, such a model can be depicted as the adversary being allowed to put probes on up to $t$ wires and to read a noisy version of the values being carried on them. In the specific case of their construction, since there isn't a circuit from which the adversary can get leakage from, the authors specify a set of variables on which noisy bit-selector functions are applied.

Q: which is their fresh re-keying scheme?

A: even if the scheme, called $\Pi_{noisy}$ in the paper, is composed of a set of algorithms, I only specify the core of GenSK, which is responsible for generating the session key. I assume the inputs are some randomness $R$ and a set of shares of the master secret key $\{msk_i\}_{i\leq d}$, which are needed in order to secure the algorithm against DPA. $$ \begin{align*} 1.\ & u_i \leftarrow R\cdot msk_i \\ 2.\ & u \leftarrow \sum_{i\leq d} u_i \\ 3.\ & sk \leftarrow H(u) \\ \end{align*} $$ The sum in the second line is computed iteratively as $((\dots (u_1 +u_2)+u_3)+\dots )+u_d$ for security reasons, while the $H$ in the third line is an hash function modelled as a random oracle. In the game, the adversary is given $sk$ and $R$ and can query leakage on a certain amount of bits of: $msk_i$, $R\cdot msk_i$ and $\sum u_i$. A further step not specified in the previous list is the application of a refreshing algorithm to the shares of the master secret key, in such a way that they look like new shares when the next session key is generated. Finally, the author prove the following: $$ \text{LPL is hard} \Rightarrow \Pi_{noisy} \text{ is secure}. $$

For many other details, the paper is available on ePrint.