Efficient Instruction Vulnerability Prediction With Heterogeneous SDC Propagation Knowledge Graph

With System on Chip integration on the rise, Silent Data Corruption (SDC) poses a significant threat to computer systems, corrupting outputs silently and without clear faults. Traditional error detection methods lack either energy efficiency or accuracy, failing to capture SDC's complex propagation patterns. To the end, in this paper, we propose a new paradigm called VP-HPKG, which leverages a Heterogeneous Program Knowledge Graph to intricately map structural interdependencies between basic blocks and instructions, enabling the exploration of potential SDC propagation paths. Specifically, first, we build an instruction execution and register fault generation system, based on which we can simulate bit-flip errors to obtain data such as program semantic information and execution status. Second, due to complex inter-instruction relations and random error propagation, we construct a multi-layer heterogeneous program knowledge graph by characterizing entities and relations of instructions, which implicates the potential path of error propagation. Then, we model contextual correlations within and among basic blocks with Graph Neural Network and Transformer, aiming to uncover abnormal inter-block jumps and extract the instruction embedding to predict vulnerable instructions within blocks accurately with low overhead. In particular, we perform 1,144,070 fault injections based on LLFI to obtain sufficient samples of SDCs. Our experimental results with 22 programs show our method outperforms the state-of-the-art baselines in terms of accuracy (average 10.3%<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\uparrow$</tex-math></inline-formula> ), F1-score (average 18.4%<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\uparrow$</tex-math></inline-formula>). In addition, the model complexity of VP-HPKG is reduced by about 2 times compared to the most competitive method, and the fault injection overhead is reduced by 30%.