This weeks study group was led by David McCann and focused on the “Compiler-based Side Channel Vulnerability Analysis and Optimized Countermeasures Application”  paper presented at the Design and Automation Conference (DAC) 2013. At a high-level, the authors consider the output for each instruction executed and determine, for each individual bit, the minimum number of key bits mixed into the result. Using this information, the compiler makes a decision (based on a threshold) on whether or not to mask the intermediate data.
Consider a toy example of the AddRoundKey operation from AES where $s = k \oplus m$ where $m$ is the input matrix, $k$ is the key matrix and $s$ is the resulting state matrix. Each bit of $s$ contains only a single bit of keyed information. The authors define security as a threshold for the minimum number of key bits affecting each intermediate state bit. The exception being an intermediate state bit that has no relation to the key bits (in this case, the security is considered infinity).
The authors incorporate the additional stages, referred to as the “security-oriented data flow analysis” (SDFA), into the LLVM compiler framework. This has some immediate advantages over applying masks directly in the source code, specifically, if the source code applies a masking scheme, and the compiler is clever enough to notice, the masking process may be identified as unnecessary computation and be optimised out. In addition to this, only the vulnerable intermediate bits are identified for masking rather than the application as a whole.
The SDFA converts the compiler intermediate-representation (IR) code to a Control-Flow Graph (CFG) where each node represents a statement in the program. The authors go on to define a set of rules for the propagation of keyed information at the input and output of the and, or, cmp, mul, div, mod, store, load and shift operations. I could define all the rules here, however this would be almost equivalent to re-writing the paper so if you are interested, please read the full text .
In the final section, the authors present experimental results with an implementation of AES-128. The results are compared to implementations presented in  and . They discuss different levels of masking and their respective code size and speed performance. The authors highlight the speed-up in performance over existing masked implementations (roughly x2.5).
It is a nice and neat presentation on how to exploit the compilers framework to identify and mask vulnerable intermediate data. I am on the fence about the speed-up results seeing as the reference implementations are optimised for 8-bit processors (whereas the target hardware was a 32-bit processor). They also state the code size increase in their work is 'acceptable' given the speed up. However there is a three fold increase in size for a first order tabulated S-Box (accredited to loop-unrolling) for a two fold speed-up. Nevertheless, a nice paper with good potential for automation of side-channel countermeasures.
Post a Comment