Science – Page 18

Uncovering AI’s Hidden Leaks: A New Approach to Privacy Auditing

15.11.2025 by qfx

$Privacy leakage is formally defined as the statistical difference, quantified by $D_{KL}(P_X||P_X')$, between the prior distribution $P_X$ over sensitive attributes and the posterior distribution $P_X'$ obtained after observing released information, thereby establishing a rigorous measure of information loss.$

Researchers are developing a formal method to identify how AI systems inadvertently reveal sensitive information through their decisions.

Decoding Stellar Chemistry Without Labels

15.11.2025 by qfx

A new deep learning approach unlocks chemical abundances from stellar spectra using unsupervised learning, paving the way for automated identification of rare stars.

Synthetic Data Gets a Privacy Boost

15.11.2025 by qfx

A new approach combines generative networks with differential privacy to create highly realistic datasets without revealing sensitive information.

Unmasking Federated Learning Attacks: Clients Can Spot the Manipulation

14.11.2025 by qfx

In linear layers, a single input’s value can be precisely reconstructed from its corresponding gradients; however, when multiple inputs activate distinct neurons concurrently, the resulting gradients become a composite weighted sum, thereby complicating the recovery process and highlighting a fundamental challenge in disentangling individual contributions within the layer.

New research reveals that active gradient inversion attacks, designed to steal data in collaborative learning, aren’t as hidden as attackers believe.

Decoding Domain Speak: Enhancing Language Models with Specialized Terminology

14.11.2025 by qfx

A new framework, TermGPT, addresses the challenges of ambiguous and sparse data in legal and financial texts to improve large language models’ understanding of specialized vocabulary.

Beyond the Numbers: Injecting Motion into Stock Market Forecasting

14.11.2025 by qfx

The KGate model, alongside a non-kinematic artificial neural network, attempts to predict Dow Jones fluctuations, but it is the kinematic-informed KIANN that demonstrates a potentially more nuanced understanding of market dynamics through its normalized input, suggesting an approach that integrates movement-based data may yield improved forecasting capabilities.

A new approach leverages principles of kinematics to refine neural network predictions, aiming for more stable and accurate long-term stock market forecasts.

Predicting What’s Next: AI for Content & Market Momentum

14.11.2025 by qfx

A new decision support system leverages artificial intelligence to forecast content virality and market growth with unprecedented accuracy.

Beyond Rational Rivals: Modeling Opponents in Complex Systems

14.11.2025 by qfx

Reinforcement learning algorithms demonstrably fall into distinct categories—policy-based and value-based—with a significant subset operating effectively without necessitating the complexities of an actor-critic architecture, as evidenced by a comprehensive taxonomy detailed in Canese et al. (2021).

A new wave of research combines machine learning with game theory to build more realistic models of strategic interaction, moving beyond assumptions of perfect rationality.

When Words Lose Meaning: AI and the Future of Job Applications

13.11.2025 by qfx

The analysis of hiring probabilities before and after the implementation of Large Language Models demonstrates a discernible shift in recruitment patterns, suggesting these models exert a measurable influence on candidate selection processes.

New research suggests that the rise of generative AI is undermining the value of written applications, creating challenges for employers and workers alike.

Responding to Crisis: The Rise of AI-Powered Emergency Systems

13.11.2025 by qfx

Diverse generative models—including Deterministic Models, Generative Adversarial Networks, and Variational Autoencoders—were rigorously compared in their ability to coordinate a swarm of four unmanned aerial vehicles, revealing varying degrees of efficacy in collective maneuvering.

New research explores how generative AI is poised to transform emergency response, enabling faster, more adaptable, and intelligent automated systems.