AI Pulse

Client-Side Probing of Deleted Ridge Statistics in Federated Unlearning

This paper addresses federated unlearning—the challenge of removing a client's data and its influence from a trained model without retraining from scratch. The authors propose a client-side probing method to verify that deleted data has been properly removed by examining ridge statistics, enabling enterprises to audit compliance with data deletion requests across distributed machine learning systems.

Privacy Failure in Split-LLM Training, The Returned Gradient Nullifies the Decoys

Researchers at arxiv.org discovered a critical privacy vulnerability in split-learning architectures where language models are distributed across multiple parties: returned gradients can be reverse-engineered to extract private training data, even when decoy data is introduced to mask sensitive information. This finding exposes a fundamental weakness in federated and split-LLM training setups commonly proposed for privacy-preserving collaborative AI development. The vulnerability is particularly relevant to regulated sectors deploying distributed LLMs across institutional boundaries.

Corporate-Family Resolution Is Not a String-Matching Problem: A Public Benchmark Stratified by Name Visibility

Researchers introduce a public benchmark for corporate-family resolution (entity matching across related organizations) that reveals name visibility significantly impacts model performance—a finding overlooked by traditional string-matching approaches. The work demonstrates that enterprise entity resolution systems require stratified evaluation beyond surface-level name similarity to handle real-world data complexity. This has direct implications for data integration and MDM (Master Data Management) in regulated sectors handling corporate hierarchies and relationships.

Learning from VAE Errors to support ECG-based Differential Diagnosis of Myocardial Scar

Researchers propose a method using Variational Autoencoder (VAE) errors to improve ECG-based diagnosis of myocardial scar, a critical indicator of heart disease. The approach leverages anomaly detection through reconstruction errors to enhance differential diagnostic accuracy in cardiology applications. This technique demonstrates potential for improving AI-assisted clinical decision support in cardiac imaging and diagnosis.

Single-Query Black-Box Calibration Auditing via Logit Bias

This paper presents a method for auditing machine learning model calibration using only a single query to a black-box model, leveraging logit bias manipulation to extract calibration information. The technique enables enterprise practitioners to verify whether model confidence scores accurately reflect true prediction probabilities without requiring access to model internals or multiple queries. This is particularly relevant for regulated industries where model reliability and transparency auditing are critical compliance requirements.

How Faithful Is Attribution for Sales Forecasting? A Counterfactual Study

This paper investigates the reliability of attribution methods—techniques used to explain AI model predictions—in the context of sales forecasting through counterfactual analysis. The research evaluates whether popular attribution approaches accurately identify which input features truly drive forecasting decisions, finding potential gaps between claimed explanations and actual model behavior. The findings are relevant for enterprises deploying interpretable AI in regulated sectors where model explainability and decision traceability are critical compliance requirements.

A Robust Watermark-based Fingerprint Framework for GNNs Ownership Verification

This paper presents a watermarking and fingerprinting framework to verify ownership of Graph Neural Networks (GNNs), addressing IP protection concerns for proprietary models. The approach embeds robust watermarks into GNN parameters that persist through model fine-tuning and extraction attacks, enabling organizations to prove legitimate ownership of their trained models.

Resilience Beyond Stationary Client Unavailability: Unlocking Efficient and Unbiased Federated Learning

This research addresses federated learning resilience when clients become unavailable during training, proposing methods to maintain model accuracy and reduce bias without relying on stationary client assumptions. The work is relevant for distributed AI systems across regulated industries where data cannot be centralized, such as healthcare networks, financial institutions, and insurance consortiums operating across multiple locations.

Federated Attack Campaign Detection via Contrastive Encoding of Threat Indicators in Gradient Updates

This paper presents a federated learning approach to detect coordinated cyber attack campaigns by analyzing gradient updates from distributed models without exposing raw threat data. The method uses contrastive encoding to identify patterns of malicious activity across organizations while preserving privacy—enabling collaborative defense without centralizing sensitive security information.

Locating and Steering Refusal Beyond Attention

This paper investigates mechanisms by which large language models refuse certain requests, finding that refusal behaviors extend beyond attention mechanisms to involve broader model components. The research presents techniques for locating and steering these refusal mechanisms, with implications for understanding model safety and control.