Louisa Crawford
September 19, 2024 10:04
NVIDIA highlights advances in AI security at Black Hat USA and DEF CON 32, emphasizing adversarial machine learning and LLM security.
According to the NVIDIA Technical Blog, NVIDIA recently demonstrated its AI security expertise at two of the most prestigious cybersecurity conferences, Black Hat USA and DEF CON 32. The events provided a platform for NVIDIA to showcase its latest advancements in AI security and share insights with the broader cybersecurity community.
NVIDIA at Black Hat USA 2024
The Black Hat USA conference is a world-renowned event that features cutting-edge security research. This year’s discussions focused on the application of generative AI tools in security and the security of AI deployments. NVIDIA’s Cybersecurity AI Lead Bartley Richardson delivered a keynote address alongside WWT CEO Jim Kavanaugh, focusing on how AI and automation are changing cybersecurity strategies.
In another session, experts from NVIDIA and partners discussed AI’s transformative impact on security posture and AI system security technologies. The AI Safety panel featured Nikki Pope, NVIDIA’s Senior Director of AI and Legal Ethics, who discussed the complexities of AI safety with practitioners. Microsoft And Google.
Daniel Rohrer, VP of Software Product Security at NVIDIA, spoke about the unique challenges of securing AI data centers in a session hosted by Trend Micro. The consensus at Black Hat was clear: deploying AI tools requires a robust approach to security, emphasizing trust boundaries and access control.
NVIDIA at DEF CON 32
DEF CON, the world’s largest hacker conference, featured a number of villages where attendees participated in live hacking challenges. NVIDIA researchers supported AI Village, hosting a popular live red team event focused on large language models (LLMs). This year’s event included a Generative Red Team challenge that led to real-time improvements to model safety guardrails.
Niki Popp gave a keynote on algorithmic fairness and safety in AI systems. The AIxCC (AIxCC), hosted by DARPA, involved red and blue teams building autonomous agents to identify and exploit code vulnerabilities. The initiative highlighted the potential of AI-based tools to accelerate security research.
Adversarial machine learning training
At Black Hat, NVIDIA and Dreadnode conducted a two-day training on machine learning (ML) that covered techniques for assessing security risks to ML models and implementing specific attacks. Topics included evasion, extraction, evaluation, reversal, poisoning, and attacks on LLM. Participants practiced executing these attacks in self-directed labs, gaining valuable hands-on experience in formulating effective defense strategies.
LLM Focus on Security
NVIDIA’s Chief Security Architect Rich Harang spoke at Black Hat about LLM security, emphasizing the importance of grounding LLM security in familiar application security frameworks. The talk focused on security issues related to the Augmented Search Generative (RAG) LLM architecture, which significantly expands the attack surface of AI models.
Attendees were advised to identify and analyze trust and security boundaries, trace data flows, and apply the principles of least privilege and minimum output to ensure strong security.
LLM Democratizing Security Assessment
At DEF CON, NVIDIA AI security researchers Leon Derczynski and Erick Galinkin introduced garak, an open-source tool for LLM security probing. Garak allows practitioners to quickly test potential LLM exploits, automating part of LLM red teaming. The tool supports about 120 unique attack probes, including XSS attacks, rapid injections, and secure jailbreaks.
Garak’s presentation and demo lab attracted a large attendance and was a significant step forward in standardizing security definitions for LLM. The tool is available on GitHub, allowing researchers and developers to quantify and compare the security of models against a variety of attacks.
summation
NVIDIA’s participation in Black Hat USA and DEF CON 32 underscores the company’s commitment to advancing AI security. The company’s contributions have provided the security community with valuable knowledge for building AI systems with a security mindset. For those interested in adversarial machine learning, NVIDIA offers self-paced online courses through the Deep Learning Institute.
For more information on NVIDIA’s ongoing efforts in AI and cybersecurity, visit the NVIDIA Technology Blog.
Image source: Shutterstock