Even the most recent developments in AI need to be improved
Analytics India Mag, October 22, 2021
GNNs are used in computer vision, NLP, and transport networks to solve various problems, but they still have a number of weak points.
GNN is a relatively new deep learning method that belongs to the category of neural networks that process data on graphs. While CNNs can only work with ordinary Euclidean data, that is, images and texts, graphs are non-Euclidean and can be used to study and analyze three-dimensional data, which GNNs can work with. In other words, they are able to mimic the reasoning process of the human brain that distinguishes them from other neural networks.
GNNs are widely used in natural language processing, computer vision, traffic networks, recommendation systems, and other areas of human activity.
However, such systems today also have a number of disadvantages. Along with graph pre training and complex graph structures, the problem of robustness also arises. GNN belongs to neural networks, and they are vulnerable to attackers. Despite the fact that such attacks do take place, it is a little more difficult to attack GNNs than other neural networks. While the attack on images and text focuses on functions, adversarial attacks on graphics require additional information.
Bleeping Computer, October 18, 2021
Credit card PINs can be guessed by a clever use of artificial intelligence technology.
Researchers have shown that it is possible to specially train a deep learning algorithm that it can guess 4-digit PIN-codes of cards in 41% of cases, even if the victim covers the entered PIN-code with his hands. To successfully carry out an attack, it is necessary to configure replicas of the target ATM, because in an attack, learning the algorithm for specific sizes and key spacing is critical. Machine learning models need to learn how to recognize keystrokes and assign specific probabilities to a set of assumptions – for this, training videos are used with people entering a PIN on an ATM.
In addition, the model can eliminate numbers based on coverage with a non-typing hand, and derives pressed numbers from movements of the other hand, estimating the topological distance between the two keys. In the end, as part of an attack using three attempts (usually this is the maximum number of attempts), the researchers recovered 5-digit PIN-codes in 30% of cases and reached 41% for 4-digit PIN-codes.
Researchers recognized such an attack as potentially dangerous and proposed a number of solutions to mitigate the risks.
Meri Talk, October 19, 2021
Pamela Isom, director of the Artificial Intelligence and Technology Office (AITO) at the Department of Energy (DoE), on Oct. 18 at the AI World virtual summit emphasized that AI users must understand the responsibilities and potential threats that AI use brings, which in turn raises the issue of organizing countermeasures.
The DoE was aiming to advance trustworthy AI and mitigate agency risks. Isom in collaboration with her colleagues released the AI Risk Management Playbook (AI RMP) available exclusively to DoE users currently.
The guide explains to users what the trusted AI lifecycle looks like and where additional security needs to be maintained by providing important guidelines for developing and deploying reliable AI including securing the supply chain of AI-powered hardware and software, providing training and testing of machine learning models, and monitoring. model output for potential security risks.
“The AI RMP is a dynamic system that offers DoE users 100+ unique risks and mitigation techniques, with the ability to expand. It also includes a search capability that allows users to filter their search according to lifecycle stage, risk type, and trustworthy AI principle,” Isom commented.
The Adversa team makes for you a weekly selection of the best research in the field of artificial intelligence security Black-box Adversarial Attacks on Network-wide Multi-step Traffic State Prediction Models ...
