Does Putin have the better hackers? Part 4 Fragile ML- architecture- gateway for cyberwarriors?

Does Putin have the better hackers? Part 4

Fragile ML- architecture- gateway for cyberwarriors?

Washington, 3/28/2022

Several media outlets reported that machine learning models used by banks and other financial institutions are uniquely exposed to cyberattacks. To that extent, they pose a significant risk. Some experts do not rule out the possibility that these models could be the target of Vergltung cyberattacks linked to Russia.

Banks‘ machine learning models currently handle or support financial services such as trading and lending. However, protecting these models from attacks is still in its infancy. That’s why U.S. President Biden had called on American companies to strengthen their cyber defenses „immediately.“ Previously, intelligence reports had come to light showing that Russia was preparing cyberattacks against the United States.

The following examples highlight where the cyber warriors could attack and what the consequences would be:

– David Van Bruwaene of Fairly AI to the „Wall Street Journal“ : Tricking the models of „over-indebted banks“ to cause huge losses „would be a kind of large-scale nuclear bomb for our economy.“

– Andrew Burt of BNJ, a law firm specializing in AI : „The vulnerabilities of AI- are significant and largely overlooked by many companies that use it- “

– Abhishek Gupta, founder of the Montreal AI Ethics Institute: the introduction of ML into software infrastructure opens up „new attack surfaces.“ The entire architecture is fragile, „like a house of cards-“

– Researcher Fabio Urbina works for Collaboration Pharmaceuticals, a North Carolina startup that specializes in AI for drug discovery.

The company has a machine-learning system called MegaSyn that can engineer molecules for therapeutic drugs to treat diseases such as Alzheimer’s.

In one experiment, Urbina tweaked the model to instead generate potentially lethal compounds ranked from high to low toxicity.

Six hours later, the AI had generated a list of 40,000 molecules. Some, though not all, of the compounds were lethal and could even be more toxic than existing chemical agents.

An article on the findings was published in !“Nature Machine Intelligence“ at a conference on biological weapons control. In it, reference is made to the potential for AI to be „dual-use,“ i.e., misusing a potentially useful system for harmful purposes.