Multilevel Heterogeneous Steganographic Container Model Implementation for Image Datasets Protection Against Poisoning Attacks

The paper considers a scenario of countering data poisoning attacks using steganographic data protection techniques. Data tampering attacks are modifications of training data, which are designed to affect the model obtained after training in a way necessary for the attacker. As a solution it is proposed to use a fragile digital watermark embedded in the data used for training and testing procedures of computer vision systems using a special container model of steganographic embedding. To verify the effectiveness of the proposed solution was performed training of neural network Yolov8n on both clean data and labeled data, the results of the study showed that the use of the proposed approach can prevent the attack of poisoning neural network training achieved by distorting the data used for training and testing.