Transport

We will structure our analysis around the AI system lifecycle.

The first stage involves collecting and preparing data for AI model training. The key data in transportation is vehicle movement data (including real-time data), cargo and freight transport data, environmental data, and infrastructure data.

The confidentiality, security, and quality of data are critical for building reliable AI systems, which is why regulatory documents place special emphasis on data supply chains [4, 5, 6]. Thorough data verification is essential and must be carried out with the involvement of industry experts [1], and in some cases, AI specialists.

The next stage is AI model creation and training. These processes must be carried out in secure software environments using trusted software. As with other critical industries, AI models developed for transportation must be free of undeclared or malicious functions—for example, those that could lead to the theft of telemetry data or changes in model behavior triggered by specific input data. Thorough analysis of a model's internal logic and its components minimizes the risk of hidden malicious payloads.

Additional requirements are placed on AI systems for vehicle control, including testing in digital twins and test sites, operational transparency, interpretability of output (decision) results for the vehicle operator (driver), seamless transfer of control to the driver, and several others. A more detailed discussion of this class of systems is provided in the section dedicated to highly automated automobile transport.

Unfortunately, high-quality training data and an accurate, functionally verified AI model do not yet fully guarantee the safety of the AI system. The fact is, during the operational phase, AI systems are vulnerable to attacks via manipulated input data. For example, the data from certain measurement systems can be deliberately modified so that the AI model's output does not match the true measurements. For vehicles using such measurement data, these attacks are particularly dangerous, as even tiny manipulations can have a significant impact on the AI system. The algorithms for attacks on input data are quite diverse, and their feasibility has been demonstrated many times in practice. One effective means of countering this type of attack is the so-called adversarial training of AI models [7, 8].

It's also worth noting the potential vulnerabilities of AI model inference equipment, but in the context of transportation, this is a difficult topic to summarize due to the huge variety of hardware solutions used. A brief description of successful attacks on AI hardware is provided in the relevant section of the website.

We have examined the process of creating an AI system and the associated risks, primarily focusing on the fields of autonomous navigation, freight transportation, and monitoring of technological processes in transport. One significant area left outside the scope of this discussion is passenger transportation, where AI can also be used to interact directly with humans.

In this field, in addition to the previously mentioned risks, there are concerns related to the improper functioning of generative AI (large language models) in AI-powered passenger service assistants. Here are just a few possible consequences of such malfunctions: