Robustness

ProTransformer robustifies transformers with a novel plug-and-play attention mechanism, significantly improving robustness across various tasks and domains without retraining.

This paper introduces a novel push-forward model for performative learning, proving the convexity of performative risk under new assumptions and linking performative learning to adversarial robustness…

Spectral algorithms for graph bisection show surprising robustness to helpful adversaries in semirandom models, with unnormalized Laplacian consistently outperforming the normalized one.

Even a few data changes can break the Benjamini-Hochberg (BH) procedure, a widely used multiple testing method, highlighting a critical vulnerability.

MALT: a novel adversarial attack, is 5x faster than AutoAttack, achieving higher success rates on CIFAR-100 and ImageNet by exploiting mesoscopic almost linearity in neural networks.

Improved facility location mechanisms are designed using ‘Mostly Approximately Correct’ predictions, exceeding prior bounds despite large prediction errors.

ELCD: The first neural network guaranteeing globally contracting dynamics!

ZORRO: A new method for learning linear models from uncertain data, providing sound over-approximations of all possible models and prediction ranges.

This work presents a computationally efficient algorithm that robustly learns a single neuron despite adversarial label noise and distributional shifts, providing provable approximation guarantees.

Zero Graves-Lai constant ensures both bounded regret and delay robustness in online decision-making, particularly for linear models.

Researchers unveil text-level graph injection attacks, revealing a new vulnerability in GNNs and highlighting the importance of text interpretability in attack success.

Strategic players can manipulate reward signals, but simple threshold policies still achieve a surprisingly good approximation to the optimal prophet value, even in this more realistic setting.

Researchers developed a novel method to inject undetectable backdoors into obfuscated neural networks and language models, even with white-box access, posing significant security risks.

Data Debiasing with Datamodels (D3M) efficiently improves machine learning model robustness by identifying and removing specific training examples that disproportionately harm minority groups’ accurac…

Boosting adversarial robustness fairness in deep neural networks, Anti-Bias Soft Label Distillation (ABSLD) adaptively adjusts soft label smoothness to reduce error gap between classes.

Proper weight initialization significantly boosts Graph Neural Network (GNN) and Deep Neural Network (DNN) robustness against adversarial attacks, highlighting a critical, often-overlooked factor.

GREAT Score: A novel framework using generative models for efficiently and accurately evaluating the global robustness of machine learning models against adversarial attacks.

FEEL-SNN enhances spiking neural network robustness by mimicking biological visual attention and adaptive leak factors, resulting in improved resilience against noise and attacks.

Speed up repeated evasion attacks on tree ensembles by 36x using feature perturbation insights!

Deep state space models (SSMs) gain adversarial robustness through an adaptive scaling mechanism, improving performance without overfitting issues.