John Laird, professor of computer science and engineering at the University of Michigan, replied: “There will be a continuous offload of mundane intellectual and physical tasks to AI and robotic systems. Nowadays, data is generally considered to be a truthful representation of the world and has been successfully used to teach AI systems to perform tasks within this world. First, adversaries can capture the physical equipment, including drones and weapon systems, in which the AI systems will live. By creating AI-based weapons and defense systems, the individual data samples used to train the models themselves become a secret that must be protected.
The military is setting a good example of this intermediate use by prioritizing the development of AI systems that increase human control, but not replace it. They can take the form of a stain or doodles on a physical target, or be hidden in the DNA of an AI system. An artificial intelligence attack (AI attack) is the intentional manipulation of an AI system with the ultimate goal of causing it to malfunction. These tests should assess the application's vulnerability to attacks, the consequences of an attack, and the availability of alternative non-AI-based methods that can be used instead of AI systems.
Once each of these questions has been sufficiently answered, they must be weighed to determine the risk posed by the system, and this should be used to make an implementation decision. These best practices should be formulated with the collaboration of security experts and experts in the field for each application, and are likely to include changes such as transmitting only data over classified or encrypted networks, encrypting the stored data to protect it even if the system is compromised, and keeping the details of the system, such as the tools and hyperparameters of the model, secret. At the same time, child labor will be reduced because robots will be able to perform tasks much cheaper and faster, forcing governments in Asia to find real solutions. In the case of inbound attacks, the adversary could then easily find attack patterns to design an attack against any system trained with the data set.
At the same time, artificial intelligence as a service, a key component of the second “shared monoculture” scenario, is also becoming more common.
