Adversarial Robustness in Machine Learning

Pervasive Backdoor Vulnerabilities in Genomic Foundation Models

This study identifies backdoor vulnerabilities in genomic foundation models through data poisoning during fine-tuning and proposes an auditing framework to detect such threats.

Safeguarding open-weight genomic foundation models through weight locking

The paper evaluates weight-locking techniques to prevent fine-tuning of genomic foundation models, finding that naive attacks fail but informed attacks can partially bypass the lock.

MedMisBench: Measuring Epistemic Resilience of LLMs Under Misleading Medical Context

The paper introduces MedMisBench, a benchmark to measure LLMs' ability to maintain correct medical judgment under misleading context, showing significant drops in accuracy when exposed to adversarial inputs.

Cerebellum-Inspired Kernel for Robust OOD Detection

This paper introduces a cerebellum-inspired kernel for OOD detection that improves performance on benchmarks and reduces computational cost.

Poisoning scientific knowledge using large language models

This paper demonstrates how large language models can generate malicious abstracts to poison scientific knowledge graphs and manipulate downstream biological applications.

Adversarial Attacks on Genotype Sequences

The paper demonstrates how adversarial attacks can corrupt genomic analyses and explores machine learning techniques to detect such manipulations.

Robust deep learning object recognition models rely on low frequency information in natural images

This study shows that robust deep learning models for object recognition rely on low spatial frequency information, as evidenced by their performance on adversarial attacks and common corruptions.

Adversarial Attacks on Protein Language Models

The paper introduces adversarial perturbations to protein sequences to disrupt 3D structure predictions, highlighting inconsistencies between mutation biological impact and model responses.

Adversarial attacks and adversarial robustness in computational pathology

This paper evaluates adversarial robustness of CNNs and ViTs in computational pathology, finding ViTs are more robust to attacks.

Adversarial manipulation of human decision-making

This paper introduces a framework for generating adversarial opponents that influence human decision-making in specific tasks, demonstrating its effectiveness through experiments on action selection and response inhibition.

Simulating a Primary Visual Cortex at the Front of CNNs Improves Robustness to Image Perturbations

A fixed-weight, biologically parameterised V1 front-end gives standard image classifiers adversarial and corruption robustness at no extra training cost; its simulated neuronal noise drives most of the adversarial gain.