Energy attack method for adaptive multi-exit neural networks

Abstract

Context:

Adaptive Multi-Exit Neural Networks (AMENNs) have emerged as a promising solution for energy-efficient and faster inference in resource-constrained environments. To ensure that these networks meet performance requirements, evaluating their energy robustness is essential. Recent works have focused on energy attacks against models in both white-box and black-box scenarios. However, existing approaches in a black-box scenarios require a significant amount of additional training data to train auxiliary models, resulting in prohibitively high costs for the attacks.

Objectives:

In this work, we leverage genetic algorithm (GA) to search for high-energy samples to conduct attacks and evaluate the energy robustness of the AMENN models directly in black-box scenario, named Energy Attack using Genetic Algorithm (EAGA).

Methods:

In the context of black-box scenarios, we propose an energy attack method based on genetic algorithm for AMENNs used in image classification tasks. By enhancing the fitness function to target high-energy consumption samples and improving population initialization and crossover mutation operations, we ensure a diverse and rich sample space for robust evaluation.

Results:

The results show that EAGA outperforms current baseline methods, demonstrating an average improvement of over 17% in the mean percentage increase in energy consumption of AMENNs. Furthermore, we guarantee the high quality of the generated attack inputs by ensuring sufficient similarity between the original image and the attack image.

Conclusion:

EAGA introduces a novel and efficient method for assessing the energy robustness of AMENNs in a black-box setting, devoid of the need for local gradient information. Through the utilization of genetic algorithms, this approach allows for a direct evaluation of model performance in resource-constrained environments. The study emphasizes the importance of EAGA in enhancing the evaluation process of AMENN models and underscores its potential to advance energy-efficient neural network deployments.