Light Backdoor Attack on Traffic Signs via Optical Triggers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing backdoor attacks for artificial intelligence security are often detectable and require direct access to the target object, making them suspicious and difficult to implement covertly.
Innovation Solution
A training method for physical light backdoor attacks involves generating a light trigger on a target object, creating a backdoor image, constructing a deep learning model, and evaluating its performance using a testing set to achieve a high attack success rate while maintaining stealthiness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If physical objects are used as triggers to improve stealthiness, then the backdoor attack becomes more covert, but direct deployment on target objects raises suspicion and reduces operational security
Solution Approach 1:
The patent uses light as an intermediary medium to deploy backdoor triggers remotely without direct physical access to target objects. The light trigger can be projected onto objects from a distance, eliminating the need for attackers to physically contact or closely approach the target, thus maintaining operational security while achieving covert deployment
Solution Approach 2:
The patent replaces traditional mechanical deployment methods (physical attachment of triggers to objects) with optical methods (projecting light patterns onto objects). This substitution allows for remote, contactless trigger deployment that is harder to detect and doesn't require direct access to the target object
2Ease of operation
If light triggers are used for remote deployment, then operational security is improved, but the complexity of generating and controlling light triggers increases
Solution Approach 1:
The patent employs a universal light trigger generation system that can create multiple different light trigger patterns using the same underlying technology platform. This multi-functional approach allows the system to adapt to different target objects and attack scenarios without requiring separate specialized equipment for each case, thereby managing complexity while maintaining versatility
3Adaptability or versatility
If more light colors (blue, green, red) are included in the light trigger, then the attack coverage is improved, but the complexity of color control and calibration increases
Solution Approach 1:
The patent controls light trigger characteristics by adjusting parameters of existing LED components rather than introducing complex new hardware. By varying the intensity and combination of blue, green, and red LED channels, the system achieves multiple color outcomes using a single multi-color LED module, simplifying the overall system while maintaining color versatility
Data Source
AI summary
Provided is a light backdoor attack method for a physical world. The method includes: generating, by a flashlight, a light trigger on a traffic sign; capturing and acquiring, by a camera, a backdoor training set, and marking all images in the backdoor training set as a target label; constructing a training set according to the backdoor training set and a clean training set; training a backdoor model using the training set to obtain a trained backdoor model; and evaluating effectiveness of light backdoor attack by applying the trained backdoor model to a testing set. The method can poison DNNs to achieve a more covert backdoor attack and provide a new idea for physical backdoor attack field.