Poison Ink: Robust and Invisible Backdoor Attack.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Neural Networks

Background:

• Deep neural networks (DNNs) are susceptible to various attacks, including adversarial, data poisoning, and backdoor attacks.
• Backdoor attacks are particularly concerning due to their stealth and ability to compromise DNNs at multiple pipeline stages.
• Existing backdoor methods often lack robustness against common data transformations and pre-processing techniques.

Purpose of the Study:

• To propose a novel backdoor attack method that is both robust and invisible.
• To overcome the limitations of existing backdoor attacks, specifically their visibility and fragility to data transformations.
• To enhance the stealthiness and robustness of backdoor attacks in deep learning.

Main Methods:

• Introduced 'Poison Ink,' a backdoor attack leveraging image structures as target areas for poison information.
• Generated inherently robust trigger patterns by utilizing image structures that retain semantic meaning during transformations.
• Employed a deep injection network to embed input-aware trigger patterns into cover images for stealthy attacks.

Main Results:

• Poison Ink demonstrated superior stealthiness and robustness compared to existing backdoor attack methods.
• The attack proved general across different datasets and network architectures, adaptable to various attack scenarios.
• Exhibited strong resistance against state-of-the-art defense techniques.

Conclusions:

• Poison Ink represents a significant advancement in creating robust and invisible backdoor attacks for deep neural networks.
• The method's resilience to data transformations and defenses makes it a potent threat in machine learning security.
• Further research into defenses against such sophisticated attacks is warranted.