Search

Progressive learning is a deep learning framework in which tasks are learned sequentially, with the capacity to leverage past knowledge from previously acquired tasks to aid in the learning and execution of new ones. It is a critical component for dealing with dynamic data, as the model is trained for a few labels and then re-used to categorize additional labels when fresh data becomes available over time. The model size is scaled for new learning while the model parameters used for previous learnings are preserved. The relevance of progressive learning is well accepted in convolutional and fully connected vanilla neural networks; however, it is primarily under-documented in a recurrent neural network for sequence classification problems. Furthermore, there is no defined strategy in the literature for tuning the recurrent layers to learn increasing tasks. We present a method for using a recurrent neural network for progressive learning. We utilize Ray Tune, a well-known hyper-parameter optimization toolbox, to standardize the methods of concurrent hyper-parameter tuning for the model that adapts with each progressive task. We demonstrate our method using two image datasets, MNIST and Devanagari.

Graph neural networks (GNNs) have shown great success in detecting intellectual property (IP) piracy and hardware Trojans (HTs). However, the machine learning community has demonstrated that GNNs are susceptible to data poisoning attacks, which result in GNNs performing abnormally on graphs with pre-defined backdoor triggers (realized using crafted subgraphs). Thus, it is imperative to ensure that the adoption of GNNs should not introduce security vulnerabilities in critical security frameworks. Existing backdoor attacks on GNNs generate random subgraphs with specific sizes/densities to act as backdoor triggers. However, for Boolean circuits, backdoor triggers cannot be randomized since the added structures should not affect the functionality of a design. We explore this threat and develop PoisonedGNN as the first backdoor attack on GNNs in the context of hardware design. We design and inject backdoor triggers into the register-transfer- or the gate-level representation of a given design without affecting the functionality to evade some GNN-based detection procedures. To demonstrate the effectiveness of PoisonedGNN, we consider two case studies: (i) Hiding HTs and (ii) IP piracy. Our experiments on TrustHub datasets demonstrate that PoisonedGNN can hide HTs and IP piracy from advanced GNN-based detection platforms with an attack success rate of up to 100%.

Logic locking has gained traction for protecting the intellectual property (IP) of integrated circuits (ICs) from untrusted foundries, test facilities, and end users. A working chip or an oracle is a significant resource for an attacker to retrieve the secret locking key. Recently, a radically different logic locking shield (DisORC) was proposed to safeguard an IC against oracle-guided attacks such as satisfiability (SAT)-based attacks which rely on scan access. This scheme was shown to be resilient on larger circuits with large sequential depths; however, it fails to protect smaller circuits, such as specialized controllers, even for large key sizes as attacks can succeed even without scan access. A truly random logic locking (TRLL) technique was proposed to thwart learning-based attacks, by making random decisions on absorbing inverters in the design into key-gates. However, small design blocks may not contain enough inverters to replace, and thereby when locked with TRLL, they may not resist learning-based attacks. Further, for large key sizes, high corruption schemes, such as TRLL, produce multiple correct keys, simplifying the key recovery for an attacker. In this work, we propose a scalable, learning-resilient, high-corruption technique to protect even small design blocks from all known oracle-less and oracle-guided attacks all the while returning only a few correct keys. Our technique 1) randomly generates enough inversions in the design to provably thwart learning-based attacks; 2) strategically inserts key-gates to return only a few correct keys and obtain high output corruption; and 3) heuristically selects key-gate locations to improve resilience against sequential SAT-based attacks. We regress our technique on 14 representative ISCAS-89 and ITC-99 benchmarks.