FeatureTrojan: Boosting stealthy and steady backdoor attacks with feature poisoning and fine-tuning injection

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Neural Networks

Background:

• Deep neural networks (DNNs) are susceptible to backdoor attacks, compromising their integrity.
• Existing backdoor attacks often use detectable triggers and exhibit unstable behaviors.
• These limitations hinder the effectiveness of DNNs in security-critical applications.

Purpose of the Study:

• To introduce FeatureTrojan, a novel backdoor attack for DNNs.
• To enhance the stealthiness and steadiness of backdoor attacks in DNNs.
• To overcome the limitations of current backdoor attack methodologies.

Main Methods:

• FeatureTrojan embeds dynamic, human-imperceptible triggers within the latent space features of clean inputs.
• It utilizes pre-trained diffusion generative models to create poisoned inputs with these triggers.
• The method fine-tunes pre-trained DNNs, injecting backdoors while maintaining parameter proximity to clean models.

Main Results:

• FeatureTrojan demonstrates significantly improved stealthiness and steadiness in backdoor attacks on DNNs.
• Experiments show an average improvement of approximately 30.0% in Attack Success Rate (ASR) against various defenses.
• The attack proves effective across multiple datasets and DNN architectures.

Conclusions:

• FeatureTrojan represents a significant advancement in DNN backdoor attack techniques.
• The proposed method offers a more robust and stealthy approach compared to existing attacks.
• FeatureTrojan's effectiveness highlights the need for more sophisticated defense mechanisms against advanced backdoor threats.