Webtagr - Technology News Summarizer

No More Adam: Learning Rate Scaling at Initialization is All You Need

2024-12-18

Researchers introduce SGD-SaI, a novel optimizer improving stochastic gradient descent. SGD-SaI addresses training imbalances by scaling learning rates at initialization for different parameter groups based on their gradient signal-to-noise ratios. Significantly more memory-efficient than AdamW, SGD-SaI matches or surpasses AdamW's performance across various Transformer-based tasks, including ImageNet classification and LLM pretraining. Its robustness and practicality are demonstrated across diverse applications, making it a compelling alternative.

(arxiv.org)

AI

Best-of-N Jailbreaking: A Novel Attack on AI Systems

2024-12-15

Researchers have developed a new AI attack algorithm called Best-of-N (BoN) Jailbreaking. This black-box algorithm repeatedly modifies prompts—randomly shuffling or capitalizing text, for example—until it elicits a harmful response from the AI system. BoN achieved impressively high attack success rates (ASRs) on closed-source language models like GPT-4o (89%) and Claude 3.5 Sonnet (78%), effectively circumventing existing defenses. Furthermore, BoN seamlessly extends to vision and audio language models, highlighting the vulnerability of even advanced AI systems to seemingly innocuous input variations. This research underscores significant security concerns in the field of AI.

(arxiv.org)

AI AI security AI attack Best-of-N Jailbreaking

Automated Assembly System Creates Cyborg Insects

2024-12-15

Scientists have developed an automated system for assembling insect-computer hybrid robots. The system uses a vision-guided robotic arm to precisely implant custom-designed bipolar electrodes onto the backs of Madagascar hissing cockroaches. The entire process takes only 68 seconds, and the assembled robots achieve steering and deceleration control comparable to manually assembled systems. A multi-agent system of 4 robots successfully navigated an obstacle course, demonstrating the feasibility of mass production and real-world applications. This research paves the way for scalable production and deployment of insect robots.

(arxiv.org)

AI insect robots automated assembly robotics

CCxTrust: A Confidential Computing Platform Leveraging Collaborative Trust from TEE and TPM

2024-12-12

CCxTrust is a novel confidential computing platform that cleverly combines the strengths of Trusted Execution Environments (TEEs) and Trusted Platform Modules (TPMs) to establish a collaborative trust framework. By leveraging the black-box Root of Trust (RoT) embedded in CPU-TEEs and the flexible white-box RoT of TPMs, CCxTrust achieves end-to-end protection of sensitive data and models, overcoming the limitations of relying on a single hardware RoT. The platform implements independent Roots of Trust for Measurement (RTM) and a collaborative Root of Trust for Report (RTR), further enhanced by a composite attestation protocol for improved security and efficiency. Experimental results demonstrate significant performance advantages.

(arxiv.org)

AI Confidential Computing Trusted Execution Environment Trusted Platform Module

Breakthrough in Reachability Analysis of the Domain Name System

2024-12-12

Researchers have presented the first decision procedure for verifying the Domain Name System (DNS), establishing its complexity as 2ExpTime. The study formalizes DNS semantics and uses a novel abstraction based on positive prefix-testable languages, reducing the DNS verification problem to the verification problem for pushdown systems. This approach effectively models attack vectors in DNS, such as amplification attacks and rewrite blackholing, providing a new theoretical foundation for ensuring DNS security and reliability.

(arxiv.org)

Development DNS security formal verification reachability analysis