Vehicle Camera Image Watermarking Against Injection Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles are vulnerable to injection attacks where false images are maliciously uploaded, leading to compromised data integrity and potential incorrect actions by the vehicle, due to existing watermarking techniques being susceptible to watermark estimation attacks.
Innovation Solution
An image watermarking system implemented by vehicle computers, involving a watermark processor that randomly selects pixel locations and generates watermarks, and a validation processor that authenticates the watermarked images using a cryptographic handshake procedure, ensuring each watermarked data segment has a unique watermark at different locations, making it robust against watermark estimation attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a fixed watermark is embedded in sensor data, then the data can be authenticated, but the watermark becomes vulnerable to estimation attacks and injection attacks
Solution Approach 1:
The patent applies dynamics by making the watermark dynamic rather than fixed. The watermark changes based on vehicle state parameters (speed, acceleration, steering angle, etc.) and sensor data characteristics. This dynamic adaptation prevents attackers from estimating or predicting the watermark pattern, as it continuously evolves with the vehicle's operational state while still enabling authentication when the vehicle is stationary or in controlled states.
Solution Approach 2:
The patent changes parameters by embedding the watermark in the frequency domain rather than the spatial domain. Specifically, the watermark is embedded in DCT (Discrete Cosine Transform) coefficients of the sensor data. This parameter transformation allows the watermark to be more robust against attacks while maintaining authentication capability, as the frequency domain representation provides different statistical properties compared to the original spatial data.
2Reliability
If watermarking is applied to all sensor data, then data integrity is protected, but the processing complexity and computational load increase
Solution Approach 1:
The patent applies local quality by selectively applying watermarking only to specific types of sensor data (image data from cameras) rather than all sensor data uniformly. Additionally, the watermark strength and embedding method can be adjusted based on the local characteristics of the data being processed, allowing for optimized computational complexity while maintaining adequate protection for the most vulnerable data streams.
Solution Approach 2:
The patent implements partial action by focusing watermarking efforts on the most critical sensor inputs (visual data from multiple cameras) rather than applying equal protection to all sensors. This selective approach reduces overall system complexity while providing adequate protection for the data streams that are most susceptible to injection attacks and have the greatest impact on vehicle operation.
3Object-affected harmful factors
If random watermarks are generated for each data segment, then resistance to estimation attacks improves, but the validation complexity increases
Solution Approach 1:
The patent implements feedback by using the vehicle's actual operational state (speed, acceleration, steering angle from the virtual driver system) as input to generate the watermark. The validation processor receives feedback about the expected vehicle state and uses this to verify the watermark's authenticity. This feedback mechanism allows the system to handle random watermarks efficiently, as the validation process leverages existing vehicle state information rather than requiring complex independent verification of random patterns.
Data Source
AI summary
A vehicle computer includes a watermark memory and a watermark processor programmed to execute instructions stored in the watermark memory. The instructions executed by the watermark processor include receiving an image captured by a camera, selecting a set of random pixel locations, generating a random watermark, and embedding the random watermark into the image at the set of random pixel locations. Another vehicle computer includes a validation memory and a validation processor programmed to execute instructions stored in the validation memory. The instructions executed by the validation processor include receiving a watermarked image, determining a random watermark, detecting an embedded watermark in the received watermarked image by selecting a set of random pixels and analyzing the selected set of random pixels for the random watermark, and authenticating the watermarked image as a result of determining that the watermarked image includes the random watermark at the set of random pixel locations.