At this week’s Virus Bulletin 2024 convention in Dublin, Eire, Sophos principal knowledge scientist Younghoo Lee will current a paper on SophosAI’s analysis into making use of massive language mannequin (LLM) synthetic intelligence to the advanced drawback of e-mail and net content material classification. Lee’s presentation will element using “multimodal” AI—utilizing fashions to look at each the textual content and pictures in e-mail and net content material to categorise the content material as benign, undesirable or dangerous.
In his paper, “Multimodal AI: The Sixth Sense for Cyber Defence,” Lee explains that adversaries can use generative AI to create convincing phishing emails and web sites, as demonstrated in SophosAI’s “Scampaign” analysis printed final yr. Moreover, conventional machine studying approaches to phishing and spam detection constructed with coaching units of current malicious and undesirable (“spam”) content material may miss new methods that don’t match current patterns.
Nonetheless, the identical expertise SophosAI examined as a possible supply of malicious e-mail and net content material will also be used to detect that type of content material, in addition to to categorise new web pages that haven’t any current popularity knowledge. By utilizing multimodal AI that appears in any respect facets of content material utilizing LLMs–fashions educated on large datasets of Web content material—to investigate e-mail headers, content material, logos and pictures, it’s doable to acknowledge new and beforehand unseen phishing emails and different malicious or undesirable content material even with out particular coaching knowledge.
Lee’s presentation is at 11:30 AM native time on Friday, October 4. A full article on the analysis will probably be printed following the presentation.
