Chaotic neural network quantization and its robustness against adversarial attacks

Abstract

Achieving robustness against adversarial attacks while maintaining high accuracy remains a critical challenge in neural networks. Parameter quantization is one of the main approaches used to compress deep neural networks to have less inference time and less storage memory size. However, quantization causes severe degradation in accuracy and consequently in model robustness. This work investigates the efficacy of stochastic quantization to enhance robustness and accuracy. Noise injection during quantization is explored to understand the impact of noise types and magnitudes on model performance. A comprehensive comparison between different applying scenarios for stochastic quantization and different noise types and magnitudes was implemented in this paper. Compared to the baseline deterministic quantization, chaotic quantization achieves a comparable accuracy, however, it achieves up to a 43% increase in accuracy against various attack scenarios. This highlights stochastic quantization as a promising defense mechanism. In addition, there is a crucial role played by the choice of noise type and magnitude in stochastic quantization. Lorenz and Henon noise distributions in stochastic quantization outperform traditional uniform and Gaussian noise in defending against attacks. A transferability analysis was discussed to understand the generalizability and effectiveness of the proposed stochastic quantization techniques. A cross-validation definition was newly evaluated in this scope to analyse the model’s stability and robustness against attacks. The study outperformed a quantization network technique and improved the model’s robustness and stability against adversarial attacks using chaotic quantization instead of deterministic quantization or even instead of stochastic quantization using traditional noise.