Research of existing approaches to embedding malicious software in artificial neural networks

Abstract

In recent years, neural networks have shown their potential as a new paradigm for solving problems in the field of information technology. They have shown their effectiveness in many areas, but training neural networks is expensive in terms of computing resources. In this regard, there are services for training networks based on cloud technologies, as well as obtaining pre-trained models. This has introduced new threats to cybersecurity. By splitting and placing malicious software to the weight parameters of the neurons of the network, it can be transmitted imperceptibly, through the channels of unreliable service providers. Seven methods of malware introduction and activation are considered, including LSB substitution (replacement of the least significant bits), Restlience training, Value-mapping, Sign-mapping, MSB reservation (saving the most significant bits), Fast substitution, Half substitution. A comparative analysis of these methods is given. Four types of triggers for software activation are considered, namely the Sign trigger, Logits trigger, Rank trigger, Fine-tuned Rank Trigger. The code with the implementation of the LSB substitution method in the Python programming language is given.