A practical design of backdoor trigger under frequency-based orthogonality constraints.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Backdoor attacks threaten deep neural network (DNN) reliability by causing malicious predictions with specific triggers.
• Existing frequency-domain attacks often adapt pixel methods or focus on stealth, facing challenges from advanced defenses.
• Current triggers lack sufficient effectiveness to reliably infect DNN models.

Purpose of the Study:

• Propose a novel frequency-domain backdoor attack paradigm using orthogonal spectrum injection.
• Mitigate spectral coupling between frequency-based triggers and benign image content for improved attack efficacy.
• Enhance both attack effectiveness and stealthiness against sophisticated backdoor defenses.

Main Methods:

• Utilize an advanced frequency filter to extract salient high-frequency regions for precise trigger injection.
• Employ a masking mechanism to suppress noise and cluster high-frequency components.
• Align phase components with a predefined trigger pattern for consistent backdoor information embedding.

Main Results:

• Demonstrate superior attack performance with triggers exhibiting higher spectral orthogonality.
• Validate the effectiveness of the orthogonal spectrum injection method through extensive experiments.
• Showcase the method's robustness against state-of-the-art backdoor defense techniques.

Conclusions:

• The proposed orthogonal spectrum injection offers a potent frequency-domain backdoor attack paradigm.
• The method effectively balances attack effectiveness and stealthiness, outperforming existing approaches.
• This research highlights the vulnerability of DNNs to advanced frequency-domain attacks and provides insights for defense strategies.