Anti-adversarial attack camera source recognition method based on locally smooth projection

Abstract

The invention discloses an anti-adversarial attack camera source recognition method based on local smooth projection, comprising the steps of: 1) camera image preprocessing; 2) constructing a camera source recognition feature extraction network; 3) generating a noise image block set; 4) Define a local smooth projection loss function; 5) Construct a front-end defense network for camera source recognition; 6) Apply a recognition model. The invention utilizes a local smooth projection to effectively suppress the adversarial noise in the feature extraction process of the camera source recognition so as to extract features with adversarial robustness, thereby realizing defense against adversarial attacks in the camera source recognition. At the same time, the present invention adopts a camera source recognition pre-defense network, which separates the feature extraction process from the defense process, which is not only easy to train, but also can be transferred to different camera source recognition networks. The invention takes into account the accuracy, robustness and transferability of the camera source recognition method based on the deep neural network.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method for identifying a camera source against adversarial attacks based on locally smooth projection. Background technique [0002] Camera source identification aims to identify the corresponding camera model by analyzing the noise in the captured image. Among many investigation and forensic issues, camera source identification has attracted great attention. In the past two years, the IEEE Signal Processing Society held the Kaggle camera source identification competition to further promote research in this direction. Camera source identification is critical for criminal investigations and trials, such as resolving copyright infringement cases and pointing out the author of illegal images. Camera source recognition also provides important evidence for other problems related to image tampering detection. The early common camera source identification method mainly used...

AI Technical Summary

Problems solved by technology

However, since deep neural networks are linear, they are very vulnerable to adversarial attacks

As long as the attacker adds tiny anti-noise to the image, the camera source recognition method based on deep neural network can generate wrong classification, which will bring a series of security problems

[0004] Since camera source recognition is different from general image classification tasks, it does not depend on image content but on image noise, general methods of defense against attacks such as noise removal are likely to be used to identify noise while removing noise while destroying images

For another type of robust optimization method, such as adversarial training, although it can defend against adversarial attacks to a certain extent, it is also very easy to reduce the accuracy of recognition.

Generally speaking, the cost of training neural networks is very expensive, and the above-mentioned robust optimization methods are difficult to be transferred to different deep neural networks.

Images