Researchers have identified a new type of security vulnerability in deep learning-based Image-to-Image (I2I) networks, termed "I2I Backdoor." This attack allows a compromised I2I network to function normally on clean inputs but produce a specific, inappropriate output when a hidden trigger is present. The researchers developed a method using universal adversarial perturbations (UAPs) as triggers and employed multi-task learning during training to enhance the attack's effectiveness. Furthermore, they demonstrated that this I2I backdoor could be extended to compromise downstream tasks like image classification and object detection by fooling models that process the backdoored images. AI
IMPACT This research highlights a novel security threat to image processing AI, potentially impacting the reliability of AI systems in sensitive applications.
RANK_REASON The cluster contains a research paper detailing a new type of attack on deep learning models. [lever_c_demoted from research: ic=1 ai=1.0]
- deep learning
- I2I Backdoor
- image classification
- Image-to-Image (I2I) networks
- multi-task learning (MTL)
- object detection
- universal adversarial perturbation (UAP)
- Wenbo Jiang
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →