SCANet: Securing the Weights With Superparamagnetic-MTJ Crossbar Array Networks

Area of Science:

• Computer Science
• Electrical Engineering
• Materials Science

Background:

• Deep neural networks (DNNs) are critical infrastructure but vulnerable to model replication attacks.
• Attacks aim to reverse-engineer synaptic weights, compromising proprietary or secret neural network designs.
• Existing defenses, like memristor-based solutions, have limitations in preventing such attacks.

Purpose of the Study:

• To propose SCANet (Superparamagnetic-MTJ Crossbar Array Networks), a novel defense mechanism against model stealing attacks on DNNs.
• To utilize the innate stochasticity of superparamagnets in magnetic tunnel junctions (s-MTJs) as a secure synapse for DNNs.
• To investigate the optimization of weight updation and power consumption through a hybrid approach of s-MTJs and conventional MTJs.

Main Methods:

• Implementing superparamagnetic magnetic tunnel junctions (s-MTJs) as synapses in DNNs.
• Utilizing the inherent stochasticity and uncontrollable switching of s-MTJs to thwart attackers.
• Proposing a modified neural network architecture to prevent replication attacks and minimize power consumption.
• Analyzing the impact of network depth and neuron count on accuracy degradation under attack.
• Evaluating SCANet's efficacy in real-time object detection scenarios.

Main Results:

• s-MTJs offer significantly higher security against model stealing compared to memristor-based solutions.
• The stochastic switching of s-MTJs provides a robust and uncontrollable defense mechanism.
• A hybrid approach using both s-MTJs and conventional MTJs allows for optimization of weight updation and power consumption.
• The study quantifies the effect of network architecture on vulnerability to attacks.
• SCANet demonstrates efficacy in real-time applications like object detection.

Conclusions:

• SCANet provides a novel and effective defense against model stealing attacks in DNNs by leveraging s-MTJ stochasticity.
• The proposed approach enhances security without compromising the functionality of DNNs.
• SCANet offers a pathway to more secure and potentially power-optimized neural network designs.
• Further research can explore the full potential of s-MTJs in securing AI systems.