📊 Статистика дайджестов

Всего дайджестов: 34022 Добавлено сегодня: 82

Последнее обновление: сегодня
📄 Fewer Weights, More Problems: A Practical Attack on LLM Pruning
2025-10-14
Авторы:

Kazuki Egashira, Robin Staab, Thibaud Gloaguen, Mark Vero, Martin Vechev

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
Model pruning, i.e., removing a subset of model weights, has become a prominent approach to reducing the memory footprint of large language models (LLMs) during inference. Notably, popular inference engines, such as vLLM, enable users to conveniently prune downloaded models before they are deployed. While the utility and efficiency of pruning methods have improved significantly, the security implications of pruning remain underexplored. In this work, for the first time, we show that modern LLM p...
ID: 2510.07985v2 cs.LG, cs.AI, cs.CR
arXiv PDF
📄 Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols
2025-10-14
Авторы:

Mikhail Terekhov, Alexander Panfilov, Daniil Dzenhaliou, Caglar Gulcehre, Maksym Andriushchenko, Ameya Prabhu, Jonas Geiping

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
AI control protocols serve as a defense mechanism to stop untrusted LLM agents from causing harm in autonomous settings. Prior work treats this as a security problem, stress testing with exploits that use the deployment context to subtly complete harmful side tasks, such as backdoor insertion. In practice, most AI control protocols are fundamentally based on LLM monitors, which can become a central point of failure. We study adaptive attacks by an untrusted model that knows the protocol and the ...
ID: 2510.09462v1 cs.LG, cs.AI, cs.CR
arXiv PDF
📄 Fewer Weights, More Problems: A Practical Attack on LLM Pruning
2025-10-11
Авторы:

Kazuki Egashira, Robin Staab, Thibaud Gloaguen, Mark Vero, Martin Vechev

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
Model pruning, i.e., removing a subset of model weights, has become a prominent approach to reducing the memory footprint of large language models (LLMs) during inference. Notably, popular inference engines, such as vLLM, enable users to conveniently prune downloaded models before they are deployed. While the utility and efficiency of pruning methods have improved significantly, the security implications of pruning remain underexplored. In this work, for the first time, we show that modern LLM p...
ID: 2510.07985v1 cs.LG, cs.AI, cs.CR
arXiv PDF
📄 Backdoor Vectors: a Task Arithmetic View on Backdoor Attacks and Defenses
2025-10-11
Авторы:

Stanisław Pawlak, Jan Dubiński, Daniel Marczak, Bartłomiej Twardowski

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
Model merging (MM) recently emerged as an effective method for combining large deep learning models. However, it poses significant security risks. Recent research shows that it is highly susceptible to backdoor attacks, which introduce a hidden trigger into a single fine-tuned model instance that allows the adversary to control the output of the final merged model at inference time. In this work, we propose a simple framework for understanding backdoor attacks by treating the attack itself as a ...
ID: 2510.08016v1 cs.LG, cs.AI, cs.CR
arXiv PDF
📄 LLMs Learn to Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions
2025-10-11
Авторы:

XuHao Hu, Peng Wang, Xiaoya Lu, Dongrui Liu, Xuanjing Huang, Jing Shao

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
Previous research has shown that LLMs finetuned on malicious or incorrect completions within narrow domains (e.g., insecure code or incorrect medical advice) can become broadly misaligned to exhibit harmful behaviors, which is called emergent misalignment. In this work, we investigate whether this phenomenon can extend beyond safety behaviors to a broader spectrum of dishonesty and deception under high-stakes scenarios (e.g., lying under pressure and deceptive behavior). To explore this, we fine...
ID: 2510.08211v1 cs.CL, cs.AI, cs.CR
arXiv PDF
📄 Leveraging Large Language Models for Cybersecurity Risk Assessment -- A Case from Forestry Cyber-Physical Systems
2025-10-10
Авторы:

Fikret Mert Gültekin, Oscar Lilja, Ranim Khojah, Rebekka Wohlrab, Marvin Damschen, Mazen Mohamad

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
In safety-critical software systems, cybersecurity activities become essential, with risk assessment being one of the most critical. In many software teams, cybersecurity experts are either entirely absent or represented by only a small number of specialists. As a result, the workload for these experts becomes high, and software engineers would need to conduct cybersecurity activities themselves. This creates a need for a tool to support cybersecurity experts and engineers in evaluating vulnerab...
ID: 2510.06343v1 cs.SE, cs.AI, cs.CR
arXiv PDF
📄 A Survey on Agentic Security: Applications, Threats and Defenses
2025-10-10
Авторы:

Asif Shahriar, Md Nafiu Rahman, Sadif Ahmed, Farig Sadeque, Md Rizwan Parvez

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
The rapid shift from passive LLMs to autonomous LLM-agents marks a new paradigm in cybersecurity. While these agents can act as powerful tools for both offensive and defensive operations, the very agentic context introduces a new class of inherent security risks. In this work we present the first holistic survey of the agentic security landscape, structuring the field around three interdependent pillars: Applications, Threats, and Defenses. We provide a comprehensive taxonomy of over 150 papers,...
ID: 2510.06445v1 cs.CL, cs.AI, cs.CR
arXiv PDF
📄 Multi-hop Deep Joint Source-Channel Coding with Deep Hash Distillation for Semantically Aligned Image Retrieval
2025-10-10
Авторы:

Didrik Bergström, Deniz Gündüz, Onur Günlü

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
We consider image transmission via deep joint source-channel coding (DeepJSCC) over multi-hop additive white Gaussian noise (AWGN) channels by training a DeepJSCC encoder-decoder pair with a pre-trained deep hash distillation (DHD) module to semantically cluster images, facilitating security-oriented applications through enhanced semantic consistency and improving the perceptual reconstruction quality. We train the DeepJSCC module to both reduce mean square error (MSE) and minimize cosine distan...
ID: 2510.06868v1 cs.IT, cs.AI, cs.CR, cs.LG, math.IT
arXiv PDF
📄 Cocoon: A System Architecture for Differentially Private Training with Correlated Noises
2025-10-10
Авторы:

Donghwan Kim, Xin Gu, Jinho Baek, Timothy Lo, Younghoon Min, Kwangsik Shin, Jongryool Kim, Jongse Park, Kiwan Maeng

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
Machine learning (ML) models memorize and leak training data, causing serious privacy issues to data owners. Training algorithms with differential privacy (DP), such as DP-SGD, have been gaining attention as a solution. However, DP-SGD adds a noise at each training iteration, which degrades the accuracy of the trained model. To improve accuracy, a new family of approaches adds carefully designed correlated noises, so that noises cancel out each other across iterations. We performed an extensive ...
ID: 2510.07304v1 cs.AR, cs.AI, cs.CR, cs.LG
arXiv PDF
📄 OptiFLIDS: Optimized Federated Learning for Energy-Efficient Intrusion Detection in IoT
2025-10-09
Авторы:

Saida Elouardi, Mohammed Jouhari, Anas Motii

Саммари на русском не найдено
Доступные поля: ['id', 'arxiv_id', 'title', 'authors', 'abstract', 'summary_ru', 'categories', 'published_date', 'created_at']
Annotation:
In critical IoT environments, such as smart homes and industrial systems, effective Intrusion Detection Systems (IDS) are essential for ensuring security. However, developing robust IDS solutions remains a significant challenge. Traditional machine learning-based IDS models typically require large datasets, but data sharing is often limited due to privacy and security concerns. Federated Learning (FL) presents a promising alternative by enabling collaborative model training without sharing raw d...
ID: 2510.05180v1 cs.LG, cs.AI, cs.CR
arXiv PDF
Показано 61 - 70 из 162 записей