WPDA: frequency-based backdoor attack with wavelet packet decomposition

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Machine Learning

Background:

• Deep neural networks (DNNs) face emerging security threats like backdoor attacks.
• Existing attacks require high poisoning ratios, making them detectable.
• Low-ratio backdoor attacks are challenging but crucial for stealth.

Purpose of the Study:

• To develop an effective backdoor attack method that operates at extremely low poisoning ratios.
• To enhance the stealthiness and resistance of backdoor attacks against defenses.
• To investigate the potential of frequency domain analysis for backdoor injection.

Main Methods:

• Proposed a frequency-based backdoor attack utilizing Wavelet Packet Decomposition (WPD).
• WPD finely decomposes images into sub-spectrograms to identify critical frequency regions for trigger insertion.
• The trigger is effectively embedded in these critical regions to form the backdoor.

Main Results:

• Achieved a 98.12% attack success rate on CIFAR-10 with a 0.004% poisoning ratio (2 out of 50,000 samples).
• The attack demonstrated exceptional effectiveness, stealthiness, and resistance.
• Successfully bypassed several advanced backdoor defenses.

Conclusions:

• The proposed frequency-based backdoor attack is highly effective even at extremely low poisoning ratios.
• Wavelet Packet Decomposition offers a promising approach for stealthy backdoor injection in DNNs.
• This method presents a significant advancement in understanding and countering backdoor vulnerabilities.