Skip the Chaos: A Self Supervised Learning-Powered Autoencoder for Adversarial Recovery

The challenges arising from the adversarial attacks possess significant threat to the Deep Learning (DL models). This is particularly in the domains of Computer Vision (CV) and image processing. In this literature a new architectural framework has been put forward for reducing the effects of the adversarial perturbations. This is carried out by utilizing an auto encoder equipped with skip connection and multi head attention mechanism. The prime objective of this study is to contribute to adversarial robustness by leveraging the idea of reconstructing the original image from an adversarial attacked image. The use of the skip connection is for preserving the crucial low-level features and utilizing the multi head attention for refining the high-level contextual information. This architectural framework has been trained with the adversarial data achieved from Projected Gradient Descent (PGD) and Fast Gradient Sign Method (FGSM) attacks using Self Supervised Learning (SSL) paradigm. For evaluating the effectiveness, we utilized the metrics like Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index Measurement (SSIM). The results achieved from this study showed that this architectural framework can successfully reconstruct the original images while maintaining the quantitative structural similarity and favorable reduction in the error scores. This study provides an architectural framework for improving the adversarial robustness having potential applications in various real-world mission critical systems.