Efficient Instruction Vulnerability Prediction With Heterogeneous SDC Propagation Knowledge Graph

With System on Chip integration on the rise, SilentData Corruption (SDC) poses a significant threat to computersystems, corrupting outputs silently and without clear faults. Tra-ditional error detection methods lack either energy efficiency oraccuracy, failing to capture SDC’s complex propagation patterns.To the end, in this paper, we propose a new paradigm calledVP-HPKG, which leverages a Heterogeneous Program KnowledgeGraph to intricately map structural interdependencies betweenbasic blocks and instructions, enabling the exploration of potentialSDC propagation paths. Specifically, first, we build an instructionexecution and register fault generation system, based on which wecan simulate bit-flip errors to obtain data such as program semanticinformation and execution status. Second, due to complex inter-instruction relations and random error propagation, we constructa multi-layer heterogeneous program knowledge graph by charac-terizing entities and relations of instructions, which implicates thepotential path of error propagation. Then, we model contextualcorrelations within and among basic blocks with Graph NeuralNetwork and Transformer, aiming to uncover abnormal inter-blockjumps and extract the instruction embedding to predict vulnerableinstructions within blocks accurately with low overhead. In partic-ular, we perform 1,144,070 fault injections based on LLFI to obtainsufficient samples of SDCs. Our experimental results with 22 pro-grams show our method outperforms the state-of-the-art baselinesin terms of accuracy (average 10.3%↑), F1-score (average 18.4%↑).In addition, the model complexity of VP-HPKG is reduced by about2 times compared to the most competitive method, and the faultinjection overhead is reduced by 30%.