28 April 2021 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
Training accurate deep learning (DL) models require large amounts of training data, significant work in labeling the data, considerable computing resources, and substantial domain expertise. In short, they are expensive to develop. Hence, protecting these models, which are valuable storehouses of intellectual properties (IP), against model stealing/cloning attacks is of paramount importance. Today's mobile processors feature Neural Processing Units (NPUs) to accelerate the execution of DL models. DL models executing on NPUs are vulnerable to hyperparameter extraction via side-channel attacks and model parameter theft via bus monitoring attacks. This paper presents a novel solution to defend against DL IP theft in NPUs during model distribution and deployment/execution via lightweight, keyed model obfuscation scheme. Unauthorized use of such models results in inaccurate classification. In addition, we present an ideal end-to-end deep learning trusted system composed of: 1) model distribution via hardware root-of-trust and public-key cryptography infrastructure (PKI) and 2) model execution via low-latency memory encryption. We demonstrate that our proposed obfuscation solution achieves IP protection objectives without requiring specialized training or sacrificing the model's accuracy. In addition, the proposed obfuscation mechanism preserves the output class distribution while degrading the model's accuracy for unauthorized parties, covering any evidence of a hacked model.
19 April 2010 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Crosstalk faults have emerged as a significant mechanism of circuit failure due to decreasing process geometries and increasing operation frequencies. Long signal nets are highly susceptible to crosstalk faults because they tend to have a higher coupling capacitance to overall capacitance ratio. Moreover, a typical long net also has multiple aggressors. In generating patterns to create maximal crosstalk induced delay on a victim net, it may be impossible to activate all aggressors logically or simultaneously to constructively induce maximum noise at the victim. Therefore, pattern generation must focus on activating a maximal subset of aggressors, weighted by actual coupling capacitance value, in close temporal proximity of the victim net transition. This max-satisfiability problem is constrained by fault effect propagation condition which involves determining an input signal assignment so as to propagate the fault effect at the victim to the primary output. In this paper, we present Automatic Test Pattern Generation (ATPG) solutions for multiple aggressor crosstalk faults for zero and unit delay models and compare the magnitude of crosstalk induced delay at the victim net. Our solution involves a combination of 0-1 Integer Linear Programming (ILP), for maximal aggressor excitation. Fault effect propagation is solved independently by using traditional stuck-at fault ATPG or by generating additional ILP constraints thus forming a integrated ILP formulation with error propagation. The effect of gate delays is summed by circuit transformation. The proposed technique was applied to ISCAS85 benchmark circuits. Results indicate that the percentage of total capacitance that can be switched varies from 75-100% for zero delay and 30-80% for variable delay case while achieving propagation of the fault effect to primary output.
30 December 2021 in IEEE Transactions on Consumer Electronics
With the continued scaling of transistor feature size, the cost of IC development has been escalating. The economics of semiconductor IC development favors high volume manufacturing, while high volume cannot be attained without developing an IC that serves many applications. Some of these applications are in low-margin Internet of Things (IoT) devices, where an SoC cannot command a high price. Consequently, without the ability to customize IC features after production, its lowest-priced application will determine an IC’s price. This motivates the manufacturers to develop chips with provisions for post-manufacturing IC customization. This paper explores how an IC can be programmed repeatedly and securely using blockchain-based smart contracts to allow users to upgrade or rent features. The availability of such a system could, for example, allow a buyer to upgrade her processor from a low-end to a high-end part by making an additional payment to the manufacturer. Implementing such a system will require remote device management capabilities. Remote device management presents unique design considerations, such as, necessity for transparency of the actions that a device takes on behalf of a user; the requirement of a trusted arbiter; and provision for management of these devices beyond the intended lifespan. To overcome the challenges of transparency, longevity, and the necessity of a trusted arbiter, we propose embedded smart contracts in concert with a blockchain. Our proposed smart contract takes the device feature configuration request as input and outputs the corresponding configuration. To support remote, secure, and authorized updates, we propose an on-die hardware module which communicates with the smart contract and enforces its functionalities. This was prototyped using a programmable system-on-chip working in concert with Ethereum blockchain. The prototype demonstrates the feasibility and practicality of the proposed solution.
30 December 2021 in IEEE Transactions on Consumer Electronics
Improving the reliability of energy distribution systems is a major concern to multiple parties as they are not only critical infrastructure themselves, but also affect other connected infrastructure. Smart Grids have been proposed to leverage the advantages of Internet of Things (IoT) to allow smarter management and faster recovery of energy distribution systems against disruptions. However, Smart Grid applications require a reliable, lightweight and fast authentication system to realize their potential in a secure manner. In this work, we propose a Strong PUF (physically unclonable function) system that can meet the stringent resource and performance constraints imposed by a Smart Grid operational environment. Our results indicate that the proposed mechanism produces a Strong PUF with close to ideal normalized uniqueness of 50 % and an accuracy of 50 % to modeling attacks using the best known machine learning algorithms. Additionally, our scheme has a low hardware overhead cost of $750~ \mu m^{2}$ in 45 nm technology and a sub-1 ms key generation time, ensuring the entire system is fast and lightweight.
13 November 2017 in IEEE Transactions on Circuits and Systems I: Regular Papers
SRAM-based weak physically unclonable functions (PUFs) have shown promise regarding tamper sensitive key storage and device ID generation. Weak PUFs rely on intrinsic process variations to produce repeatable and unique start-up behavior. However, noise in the system can affect the start-up behavior and introduce errors. A number of solutions, such as fuzzy extraction and error correcting codes have been proposed to alleviate the effect of noise and generate stable keys. Unfortunately, the overhead from these techniques grows superlinearly with increasing error rate. Alternatively, accelerating device aging or burn-in has been shown to reduce start-up error rate significantly, which leads to reduced overhead for error correction. Unfortunately, burn-in is a time-intensive process and accrues significant production cost. This paper proposes a method to reduce the cumulative burn-in time via quantification of the minimum burn-in requirements for chip. We propose a low-cost proxy to measure the degree of process variation of each device at birth and use previously proposed device aging models to determine the burn-in requirements. Our results show that this procedure reduces cumulative burn-in cost without compromising the resultant reliability of weak PUFs. Furthermore, we analyze the effect of having additional PUF cells, alternate weak PUF designs and transistor technologies on burn-in requirements.