AI Sec Watch

This research paper describes a new attack called Knowledge Transfer Attack (KTA) that can steal private labels (the correct answers or classifications) from graph-based vertical federated learning (GVFL, a system where multiple parties collaborate on machine learning while keeping their data private). Unlike previous attacks that required unrealistic access to training data or labeled examples, KTA only needs auxiliary graphs from unrelated domains to infer the private labels, making it a more practical threat to real-world GVFL systems.

Website fingerprinting (WF) attacks identify which websites users visit on Tor, a privacy network, but struggle when traffic patterns differ between training and real-world scenarios. This research presents UDA-WF, a new method using unsupervised domain adaptation (a machine learning technique that helps models work across different data distributions) to identify websites more efficiently with less training data. UDA-WF reduces the auxiliary data needed by 95% while maintaining 97.37% accuracy.

Backdoor attacks (hidden triggers that manipulate AI model predictions while keeping normal performance intact) are a serious security threat to deep neural networks (machine learning models with many layers). This paper presents PVDI, a defense method that removes backdoors by selectively preserving important attention patterns (the AI's focus on relevant input features) while disrupting irrelevant ones, successfully reducing attack success rates without hurting the model's normal performance.

ParaVul is a framework that uses LLMs (large language models, AI systems trained on huge amounts of text) combined with RAG (retrieval-augmented generation, where an AI pulls in external documents to answer questions) to detect vulnerabilities in smart contracts (self-executing programs on blockchain networks). The framework improves detection accuracy and reduces computational costs by using a new fine-tuning technique called SLoRA (Sparse Low-Rank Adaptation) and combining multiple detection methods through a verification module.

DiffHash is a new attack method that can fool deep hashing models (AI systems used to quickly search through large collections of images) by generating adversarial examples (slightly modified images that trick AI systems). Unlike older attacks that simply change individual pixels, DiffHash uses text descriptions from a Large Language Model to guide the creation of fake images that appear normal to humans but cause the hashing model to return wrong search results for a target image.

This research addresses a security weakness in neural steganography (hiding secret messages inside audio files using AI networks), where sender and receiver models must stay perfectly synchronized, creating risks of information leakage. The researchers propose a decoupled framework based on the destruction-restoration principle, where embedding works through a destructive operation and recovery uses a separate neural network, allowing the sender to change their embedding network without breaking the receiver's ability to extract the hidden message.

Network traffic patterns constantly change, causing traditional malicious traffic detection systems to become less effective over time, a problem called concept drift (when the patterns an AI learned on no longer match real-world data). Researchers developed Argus, a framework that automatically detects when traffic patterns shift, identifies new malicious patterns without human help, and continuously updates itself to maintain high detection accuracy even as attacks evolve.

Face-swapping deepfakes (AI-generated videos or images where one person's face is replaced with another) are widely misused for fraud and misinformation, and while detection tools exist, there has been little work on tracing and recovering the original face that was replaced. This paper presents FaceReclaim, a new AI method that uses diffusion models (neural networks trained to gradually refine noisy images into clear ones) to restore the original face from a deepfaked image by separating facial attributes like expressions from identity information.

According to Stanford economist Mordecai Kurz, tech billionaires are concentrating technological power in a way that weakens democracy, similar to patterns seen during industrialization and the first Gilded Age. Kurz argues that tech moguls often see themselves as superior beings meant to reshape society, citing examples like Anthropic's CEO claiming AI could become a transcendent good while potentially causing mass unemployment.