Luisa Crawford
Sep 19, 2024 10:04
NVIDIA highlights AI safety developments at Black Hat USA and DEF CON 32, emphasizing adversarial machine studying and LLM safety.
NVIDIA lately demonstrated its AI safety experience at two of essentially the most prestigious cybersecurity conferences, Black Hat USA and DEF CON 32, in response to the NVIDIA Technical Weblog. The occasions supplied a platform for NVIDIA to showcase its newest developments in AI safety and share insights with the broader cybersecurity group.
NVIDIA at Black Hat USA 2024
The Black Hat USA convention is a globally acknowledged occasion that options cutting-edge safety analysis. This 12 months, discussions highlighted the functions of generative AI instruments in safety and the safety of AI deployments. Bartley Richardson, NVIDIA’s Director of Cybersecurity AI, delivered a keynote alongside WWT CEO Jim Kavanaugh, specializing in how AI and automation are remodeling cybersecurity methods.
Different classes featured specialists from NVIDIA and its companions discussing the revolutionary affect of AI on safety postures and strategies for securing AI programs. A panel on AI Security included Nikki Pope, NVIDIA’s Senior Director of AI and Authorized Ethics, who mentioned the complexities of AI security with practitioners from Microsoft and Google.
Daniel Rohrer, NVIDIA’s VP of Software program Product Safety, addressed the distinctive challenges of securing AI knowledge facilities in a session hosted by Pattern Micro. The consensus at Black Hat was clear: deploying AI instruments necessitates a strong method to safety, emphasizing belief boundaries and entry controls.
NVIDIA at DEF CON 32
DEF CON, the world’s largest hacker convention, featured quite a few villages the place attendees engaged in real-time hacking challenges. NVIDIA researchers supported the AI Village, internet hosting well-liked dwell red-teaming occasions centered on giant language fashions (LLMs). This 12 months’s occasions included a Generative Pink Staff problem, which led to real-time enhancements in mannequin security guardrails.
Nikki Pope delivered a keynote on algorithmic equity and security in AI programs. The AI Cyber Problem (AIxCC), hosted by DARPA, noticed purple and blue groups constructing autonomous brokers to determine and exploit code vulnerabilities. This initiative underscored the potential of AI-powered instruments to speed up safety analysis.
Adversarial Machine Studying Coaching
At Black Hat, NVIDIA and Dreadnode performed a two-day coaching on machine studying (ML), masking strategies to evaluate safety dangers in opposition to ML fashions and implement particular assaults. Subjects included evasion, extraction, assessments, inversion, poisoning, and assaults on LLMs. Contributors practiced executing these assaults in self-paced labs, gaining hands-on expertise important for shaping efficient defensive methods.
Deal with LLM Safety
NVIDIA Principal Safety Architect Wealthy Harang introduced on LLM safety at Black Hat, emphasizing the significance of grounding LLM safety in a well-recognized software safety framework. The speak centered on the safety points related to retrieval-augmented technology (RAG) LLM architectures, which considerably increase the assault floor of AI fashions.
Attendees have been suggested to determine and analyze belief and safety boundaries, hint knowledge flows, and apply the ideas of least privilege and output minimization to make sure strong safety.
Democratizing LLM Safety Assessments
At DEF CON, NVIDIA AI Safety Researchers Leon Derczynski and Erick Galinkin launched garak, an open-source instrument for LLM safety probing. Garak permits practitioners to check potential LLM exploits rapidly, automating a portion of LLM red-teaming. The instrument helps practically 120 distinctive assault probes, together with XSS assaults, immediate injection, and security jailbreaks.
Garak’s presentation and demo lab have been well-attended, marking a big step ahead in standardizing safety definitions for LLMs. The instrument is on the market on GitHub, enabling researchers and builders to quantify and examine mannequin safety in opposition to varied assaults.
Abstract
NVIDIA’s participation in Black Hat USA and DEF CON 32 highlighted its dedication to advancing AI safety. The corporate’s contributions supplied the safety group with invaluable information for deploying AI programs with a safety mindset. For these focused on adversarial machine studying, NVIDIA presents a self-paced on-line course via its Deep Studying Institute.
For extra insights into NVIDIA’s ongoing work in AI and cybersecurity, go to the NVIDIA Technical Weblog.
Picture supply: Shutterstock
