ATTRITION: Attacking Static Hardware Trojan Detection Techniques Using Reinforcement Learning

Hardware Trojan Insertion Using Reinforcement Learning

“…Also, the achieved performance is not stable [2]. Implementation of the method is neither trivial nor available online for researchers [22].…”

“…Gohil et al [22] proposed ATTRITION, another RL-based HT insertion platform where signal probability is the target upon which the trigger nets are selected. The agent tries to find a set of so-called compatible rare nets, i.e., a group of rare nets that can be activated together with an input test vector.…”

“…The reward function, which is a combination of the number of activated nets and their SCOAP [24] parameters, pushes the agent to activate as many signals as possible. Despite its higher HT detection rate than MERO and TARMAC, the algorithm was not tested on any HT benchmarks [22].…”

“…Lyu and Mishra introduced TARMAC [17], which maps the trigger activation problem to the clique cover problem and uses a SAT-solver to generate test vectors for each maximal satisfiable clique. Though it performs well, it lacks scalability and stability, and the implementation is not readily available for researchers [6], [11].…”

“…We take "The Seeker's Dilemma" approach in creating realistic HT benchmarks to address the shortcomings of existing HT benchmarks. The HT location and size are known to the security researcher in current benchmarks such as [9]- [11]. This enables the defense side to fine-tune their HT detectors to showcase strong HT detection rates.…”