Adversarial Robustness in Machine Learning
New papers in adversarial robustness in machine learning, the code and data they release, and the open problems their authors admit. Rebuilt every day from OpenAlex and arXiv.
New papers
Two-for-One: Image-based Transformations to Mitigate Adversarial Attacks and Random Faults at Once in Safety-critical AI Systems
In this article, we introduce enhanced DMR and TMR strategies for image-based object detection, leveraging image transformations during inference to help reduce the impact of adversarial inputs, while preserving the inherent advantages of diverse redundancy for safety purposes.
TP-Net: Self-Supervised Pretrained Network Traffic Representations for Few-Shot Threat Detection in IoT and Sensor Networks
First, we build a protocol-fair benchmark that evaluates six supervised-traffic baselines and TP-Net under the same per-episode linear head with a frozen encoder, showing that TP-Net’s SSL encoder stays competitive, trailing the strongest fully trained supervised model by at most 9.75 pp,…
Feature-Decoupled Contrastive Self-Distillation for Black-Box Watermark Removal
To address this problem, this paper proposes a feature-decoupled contrastive self-distillation method for watermark removal against black-box-verifiable watermarks under an editable-model threat model.
Who Wins the Window? Agentic Vulnerability-Response Benchmarks Measure Model Competence, Not Vulnerability Difficulty
A measurement-validity study of agentic CVE benchmarks.
Mind the Attack: How security threats are shaping software deployed on Unmanned Aerial Vehicles
Context Autonomous unmanned aerial vehicles (UAVs) are increasingly used in applications that depend on reliable software operation, making security an important concern.
Phase-Locked LLM Inference and Swarm Coherence for Autonomous UAVs under Electronic Warfare via Rössler Dynamics and Dynamic F-LoRA
We introduce Phase-Locked DynamicSampling (PLDS) and Dynamic F-LoRA to maintain LLM stability and execution efficiencyon resource-constrained embedded systems.
Fake Wi-Fi (Evil Twin) Detection Using Machine Learning: A Review and a Random Forest-Based Detection Framework
Public Wi-Fi hotspots, although convenient, compromise the privacy and security of the users since attackers can launch Evil Twin attacks by deploying rogue access points that mimic the legitimate ones and lure devices to connect to them.
Deepfake Detection in Real-Time Adversarial Environments
Generative models have evolved from an offline media issue into an operational challenge that requires real-time responses to deepfakes.
Development of an Image-Translating Model to Counter Adversarial Attacks
Deep neural networks, including the classifiers that underpin deepfake detection, are vulnerable to adversarial examples: inputs that are almost indistinguishable from genuine data yet are labeled incorrectly by the network.
AN EARLY WARNING LAYER FOR AUTONOMOUS AI AGENTS: A TRAINING-FREE SIGNAL BASED ON LAG-1 AUTOCORRELATION
In June 2026, an AI agent built by OpenAI reached into an Australian government health data portal while doing a research task.
A Hybrid Passive–Active Framework for Server-Side Request Forgery Detection Using Machine Learning
We present ML-SSRFinder, a dual-mode framework that unifies three analysers.
Impact of DDoS Attacks and Mitigation Strategies in Cloud Applications
This resource presents the research findings and literature review for the paper titled "Impact of DDoS Attacks and Mitigation Strategies in Cloud Applications" by Rabin Kumar Chhatkuli (2026).
GuardiaMCP: A Free, Offline, Test-Verified Defensive Layer for the Model Context Protocol
The Model Context Protocol (MCP) lets AI assistants such as Claude, Cursor and VS Code call external "servers" that can read local files, use the user's credentials and inject instructions into the model's context.
The Existence of Implicit Knowledge Graphs Conditional Construction and Analysis Based on Internal Transformer Representations
Implicit knowledge graphs describe knowledge organization within large models as structures of semantic relations that emerge in response to an input.
Prompt-Injection Defences on a Small Local Agent: What Works, What It Costs, and How Optimistic Development Numbers Are
An AI agent that reads email, files and web pages for a user can be hijacked by instructions an attacker plants in that data: indirect prompt injection.
X-FCA: Diagnosing Explanatory Decoupling in Financial Information Systems via Counterfactual XAI and Conformal Prediction
Automated financial information systems increasingly delegate sentiment-drivendecisions to transformer-based language models, yet whether those decisionsrest on semantic content or on representational shortcuts is invisible to stan-dard accuracy metrics.
Towards Explainable and Trustworthy Traffic Sign Recognition for Safe Autonomous Driving Via Neurosymbolic AI
As autonomous vehicles become increasingly integrated into modern transportation systems, the need for robust, reliable, and interpretable perception systems has become critical.
Stateful Monitor-Router-Rollback Regulation for Robust Functional Recovery in Large Language Model Agents
Prompt injection can divert large language model (LLM) agents from authorized objectives and contaminate subsequent conversational context.
Developing a machine learning-based algorithm for secure routing in autonomous vehicles
Over the last several years, we have noticed that the administration and car companies have been refraining from investing capital in connected autonomous cars.
The Defendability Bounds: Architectural Conditions of Governability under Generative Adversaries
The series' deposited papers locate the frontier of detection: the content channel collapses operationally under budgeted mimicry [Rouxel 2026a, Hypothesis 3.12, Theorem 3.13], the temporal escape closes [Rouxel 2026b, Corollary 4.2, Corollary 4.4], durable resistance rests on three…
Released code and data
TP-Net: Self-Supervised Pretrained Network Traffic Representations for Few-Shot Threat Detection in IoT and Sensor Networks
First, we build a protocol-fair benchmark that evaluates six supervised-traffic baselines and TP-Net under the same per-episode linear head with a frozen encoder, showing that TP-Net’s SSL encoder stays competitive, trailing the strongest fully trained supervised model by at most 9.75 pp, without any encoder fine-tuning while supervised baselines require retraining.
Dataset Explanation Signatures for Visual Recognition: A Model-Free Region-Level Reference for Shortcut and Label-Quality Auditing
We present Dataset Explanation Signatures (DES) for images: a region-level reference computed from a labelled image set alone, with no fitted predictor.
Valid but Wrong: Silent Enum Substitution in Tool Calls, and Why Schema Validation Cannot See It
The items, prompts and all model outputs are in the repository: https://github.com/LakshitF/valid-but-wrong
First Line or Last Word: Where a Stored Order Sits in a Product Review and Whether a Web LLM Assistant Obeys It
We release payloads, transcripts, judge labels, and the counting scripts.
Systematic Evaluation of Adversarial Jailbreaking Attacks and Defenses Across Seven Large Language Models
Our code and datasets are publicly available at https://github.com/Mdniloykhan/jailbreak-eval-7llms.
Difficulty Calibration Across Time for Membership Inference in Diffusion Models
Code is available at https://github.com/MichaelJoseph777/TSC.
Elucidating the Design Space of LM Data Attribution
We introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which we fine-tune LLMs on entirely fictitious knowledge, providing unambiguous ground truth for attribution.
Appreciation Teaches Unauthorized Execution: Self-Evolution Degrades the Robustness of Personal Agents to Indirect Prompt Injection
The code is available at https://github.com/SaFo-Lab/Evolve-IPI
Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection
Code and data partitions are available for reproducibility.
AegisFlow: Training-Free Non-Myopic Path-Safe Guided Flow Matching
Code is available at https://github.com/LiuKunpeng03/AegisFlow.
Open problems authors admit
Mind the Attack: How security threats are shaping software deployed on Unmanned Aerial Vehicles
Future research should prioritize integrated cyber-physical defenses, closed-loop evaluation, reproducibility, collaboration, and resource-aware deployment.
GuardiaMCP: A Free, Offline, Test-Verified Defensive Layer for the Model Context Protocol
We describe the threat model, the design, the detection approach, and an honest account of what the tool does not cover.
A Review of Out-of-Distribution Generalization in Text Classification Models
Finally, it identifies the limitations of existing studies and proposes future research directions.
From LLMs to MLLMs to Agents: A Survey of Emerging Paradigms in Jailbreak Attacks and Defenses within LLM Ecosystem
To address these limitations, we provide an updated synthesis of recent work and introduce future research directions in areas such as dataset construction, evaluation framework optimization, and strategy generalization.
Adversarial Process Disturbances at a Deployable Alarm Budget: Detector Evaluation on a Simulator and a Real Plant
Three conditions prior work does not cover: disturbances authored against a threat model rather than native faults; one construction applied to both processes; and detection read at a declared false-alarm budget. The drop persists throughout.
ADVERSARIAL ATTACKS ON FRAUD DETECTION MODELS: A REVIEW OF EVASION, DATA POISONING, AND CAMOUFLAGED FRAUDSTERS
Future work should test defenses under attack and report results in a standard format.
On the Formalization of Constrained Adversarial Attacks in Network Traffic Analysis
We conclude that domain constraints must be accounted for as comprehensively as possible while preserving practical applicability, and that the effectiveness of constrained attack methods and constraint-handling approaches remains an open question.
Advanced machine learning for space cybersecurity
Finally, the discussion highlights future research directions such as physics-informed learning, neuromorphic computing, post-quantum security, and governance frameworks needed for responsible deployment of ML-based cybersecurity in space networks.
uTHCD-C: A Corruption-Robustness Benchmark for Tamil Handwritten Character Recognition
By providing standardized corruption configurations, evaluation protocols, and comparative results, uTHCD-C establishes a reproducible basis for future research on robust and computationally efficient Tamil HCR.
An Operational Scorecard for Evaluating Enterprise AI Agents on Data Platforms: A Reproducible Test Protocol for Reliability, Governance, Security, Recovery, and Cost
Numerical examples are hypothetical, and implementation and independent validation remain future work.