A noise-assistant network for tampering detection via inconspicuous feature enhancement and multi-perspective perception

In response to malicious tampering with digital images, neural networks are employed to detect tampering, thereby enhancing digital information security. The effectiveness of neural networks in tampering detection is profoundly influenced by the optimal utilization of fingerprint features within altered images. To enhance conspicuous noise remnants in forgery images, we propose a Noise-Assistant Network. The model acquires noise feature blocks in the feature extraction module containing FusionConv2D. The enhanced noise feature kernel is then employed to activate tampered feature representation in high-dimensional space. This activation takes place in both the feature enhancement module and the multi-perspective perception module, designed for the coarse and fine classification phases of tampering localization, respectively. Unlike introducing noise solely once within the neural network, our methodology involves introducing tampered noise information at distinct stages, achieving a cumulative enhancement effect. we create a synthetic tampering dataset, Syn-Pairs Dataset, containing positive and negative samples to amplify differences between tampered and non-tampered regions with similar semantic content. We use this dataset in the pre-training of the Noise-Assistant Network. Furthermore, the experiment using CASIA, COLUMBIA, NIST16, COVERAGE, DSO-1 and IMD databases yield outstanding results across various evaluation metrics, including AP series and F1. Notably, the AP₅₀ for DSO-1 reaches an impressive value of 0.867, indicating high performance in tampering detection.