Gartner has announced the top technology trends for 2023 and AI Trust, Risk and Security Management (AI TRiSM) is one of them.
Gartner’s list of 10 positions should be used to analyze the potential impact of these trends on strategies already developed and in development, including revenue growth or digital acceleration. Thus, new trends may represent a risk or, conversely, new opportunities for organizations.
AI TriSM aims to maintain the governance of the AI model, its reliability, security, data protection, robustness, and efficiency. It contains appropriate methods for explaining results, controlling privacy and ethics, managing the security of an AI model, combined with methods for quickly deploying the model.
Read about TRiSM and other technology trends in the full article.
Tech Xplore, October 18, 2022
Ingrid Fadelli
The deep learning-based malware detector based on new and promising transformer model is also vulnerable to adversarial attacks, according to researchers from the College of Engineering in Pune, India. They conducted a study focusing on the transformer-based detector, a class of deep learning models that can weigh different parts of the input differently.
Attackers are constantly coming up with new and complicated methods to achieve their goals, be it stealing users’ confidential information or damaging computer systems. In their turn, scientists and researchers are trying to generate new effective methods for detecting and preventing cyber attacks.
Currently, most malware detectors are based on machine learning, which leaves security vulnerabilities in their application. Using an adversarial attack, such as data perturbation or editing, the machine learning algorithm can misclassify the data and may identify malware as something safe.
Researchers Jakhotiya, Patil, and Rawlani have developed their own malware detection system containing three main components: assembly module, static function module, and neural network module.
Read the full article following the link and find out protection strategies the researchers offer and conclusions they came to.
Hackaday, October 20, 2022
Dan Maloney
Zuxuan Wu, Ser-Nam Lim, Larry S. Davis, and Tom Goldstein from University of Maryland have conducted research with the idea to inform humans on the imperfectness and vulnerabilities of vision systems.
A YOLO-based vision system was used in this case, standing for “You Only Look Once” that refers to an algorithm that detects and recognizes various objects in a picture in real-time.
The researchers generated an adversarial pattern and a large set of training images, some of which contain people. As a result, this pattern can prevent humans from being detected and recognized. Those attacks are not new and Adversa AI Research team had previously demonstrated how it can be done in physical world but it certainly a great example for increasing awareness in this topic.
Schneier, October 21, 2022
The security of ML systems is becoming increasingly complicated because of the variety of attacks on it.
Two researchers Eugene Bagdasaryan and Vitaly Shmatikov from Cornell Tech studied a novel threat to neural sequence-to-sequence (seq2seq) models. Training-time attacks cause models to “spin” their outputs so as to support a certain sentiment or point of view. It happens when the input contains trigger words an adversary has chosen, say, the name of some individual or organization.
Model spinning enables propaganda-as-a-service. An attacker can create customized language models that produce desired spins for triggers, then deploy these models to generate disinformation, or else inject them into ML training pipelines.
The researchers developed a new backdooring technique and evaluated the attack on language generation, summarization, and translation models with different triggers and meta-tasks like sentiment, toxicity, and entailment.
Read the full research following the link.
Stay up to date with what is happening! Get a first look at news, noteworthy research and worst attacks on AI delivered right in your inbox.
Boston Dynamics and five other robot makers pledge not to weaponize their robots Interesting Engineering, October 9, 2022 Recently, the largest robot manufacturers agreed and signed a commitment not to ...
