Boosting the transferability of adversarial examples via stochastic serial attack.

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are susceptible to adversarial examples, which are subtly modified inputs designed to cause misclassification.
• Adversarial examples exhibit transferability, meaning examples crafted for one DNN can fool others, raising concerns about transfer-based attacks.
• Current transfer-based attacks often use a single local model, leading to poor transferability due to overfitting, or ensemble methods with high memory costs.

Purpose of the Study:

• To propose a novel attack strategy, the stochastic serial attack (SSA), to enhance the transferability of adversarial examples against DNNs.
• To reduce the memory consumption associated with generating adversarial examples compared to existing methods.
• To ensure adversarial examples do not overfit to specific local models.

Main Methods:

• The proposed stochastic serial attack (SSA) employs a serial strategy for attacking local models, contrasting with parallel approaches.
• Local models are stochastically selected from a large set during the attack process.
• This approach aims to mitigate overfitting to individual local model weaknesses.

Main Results:

• SSA demonstrated effectiveness in improving the transferability of adversarial examples.
• The method significantly reduced memory consumption during the adversarial example generation process.
• Experiments on ImageNet and the NeurIPS 2017 adversarial competition dataset validated SSA's performance.

Conclusions:

• The stochastic serial attack (SSA) offers an effective solution for generating highly transferable adversarial examples.
• SSA provides a memory-efficient alternative to existing ensemble attack methods.
• The findings highlight the potential of SSA for robust adversarial attack research and defense.