Related Experiment Videos
Perceptual image hashing via feature points: performance evaluation and tradeoffs
1Xerox Innovation Group, El Segundo, CA 90245, USA. vishal.monga@xeroxlabs.com
Summary
We developed a new image hashing method using key visual features that resist common distortions and attacks. This robust algorithm accurately detects malicious image manipulations while maintaining data integrity.
Area of Science:
- Computer Vision
- Digital Image Processing
- Cryptography
Background:
- Image hashing is crucial for content authentication and copyright protection.
- Existing methods often struggle with perceptual distortions and adversarial attacks.
- Robustness against common image manipulations is a significant challenge.
Purpose of the Study:
- To propose a novel image hashing paradigm resilient to perceptual distortions and adversarial manipulations.
- To develop an iterative feature detector for extracting geometry-preserving feature points.
- To enhance hash algorithm security through probabilistic quantization.
Main Methods:
- Utilizing visually significant feature points invariant to perceptually insignificant distortions.
- Employing an iterative feature detector for robust feature extraction.
- Applying probabilistic quantization to derived features for enhanced security.
Main Results:
- The proposed image hashing algorithm demonstrates robustness against standard benchmark attacks (e.g., Stirmark).
- It successfully withstands compression, scaling, small-angle rotation, and common signal-processing operations.
- Accurate detection of content-changing (malicious) manipulations was achieved.
- Receiver operating characteristic (ROC) curves confirm perceptual robustness and low misclassification rates.
Conclusions:
- The developed image hashing scheme offers a reliable solution for image authentication.
- It effectively balances perceptual robustness with resistance to adversarial attacks.
- The method shows significant promise for secure digital image management.