A Watermarking-Based Framework for Protecting Deep Image Classifiers Against Adversarial Attacks

Abstract

Although deep learning-based models have achieved tremendous success in image-related tasks, they are known to be vulnerable to adversarial examples---inputs with imperceptible, but subtly crafted perturbation which fool the models to produce incorrect outputs. To distinguish adversarial examples from benign images, in this thesis, we propose a novel watermarking-based framework for protecting deep image classifiers against adversarial attacks. The proposed framework consists of a watermark encoder, a possible adversary, and a detector followed by a deep image classifier to be protected.

At the watermark encoder, an original benign image is watermarked with a secret key by embedding confidential watermark bits into selected DCT coefficients of the original image in JPEG format. The watermarked image may then go through possible adversarial attacks. Upon receiving a watermarked and possibly attacked image, the detector accepts it as a benign image and passes it to the subsequent classifier if the embedded watermark bits can be recovered with high precision, and otherwise rejects it as an adversarial example. The embedded watermark is further required to be imperceptible and robust to JPEG re-compression with a pre-defined quality threshold.

Specific methods of watermarking and detection are also presented. It is shown by experiment on a subset of ImageNet validation dataset that the proposed framework along with the presented methods of watermarking and detection is effective against a wide range of advanced attacks (static and adaptive), achieving a near zero (effective) false negative rate for FGSM and PGD attacks (static and adaptive) with the guaranteed zero false positive rate. In addition, for all tested deep image classifiers (ResNet50V2, MobileNetV2, InceptionV3), the impact of watermarking on classification accuracy is insignificant with, on average, 0.63% and 0.49% degradation in top 1 and top 5 accuracy, respectively.