Robust and Transferable Backdoor Attacks Against Deep Image Compression With Selective Frequency Prior.

Area of Science:

• Computer Vision
• Machine Learning Security
• Image Compression

Background:

• Deep learning models excel in image compression but are vulnerable to backdoor attacks.
• Backdoor attacks use trigger patterns to manipulate model behavior.

Purpose of the Study:

• To propose a novel multi-trigger backdoor attack against learned image compression models.
• To demonstrate the attack's effectiveness in degrading compression quality and impacting downstream tasks.
• To enhance attack robustness and transferability.

Main Methods:

• Developed a frequency-based trigger injection model using the Discrete Cosine Transform (DCT) domain.
• Designed dynamic loss functions for efficient training and optimized attack objectives.
• Implemented a two-stage training schedule with robust frequency selection for enhanced resistance.
• Incorporated classification boundary shifting for improved cross-model and cross-domain transferability.

Main Results:

• Successfully injected multiple backdoors with corresponding triggers into a single image compression model.
• Demonstrated attack effectiveness against compression quality (bit-rate, reconstruction quality).
• Showcased attack impact on downstream computer vision tasks like face recognition and semantic segmentation.
• Validated attack resistance against defensive pre-processing methods.

Conclusions:

• The proposed frequency-based backdoor attack is effective against learned image compression models.
• The attack can compromise both compression performance and downstream task accuracy.
• The method shows robustness and transferability, posing a significant security threat.