Silent Sabotage: A New Threat to Federated Learning
Researchers have uncovered a sophisticated attack that exploits attention mechanisms in federated self-supervised learning to inject hidden backdoors into models.
Science
Entanglement’s Boundaries: Mapping Entropy in Dynamic Spacetimes
![Entanglement wedges remain distinct, as demonstrated by the absence of multiple entry of any given HRT surface-represented by a red curve on a Cauchy slice-into another, while adherence to the condition that the intersection of the Cauchy slice with the boundary of spacetime-[latex]\Sigma \cap \partial M[/latex]-contains the spacelike boundaries [latex]\partial B[/latex] and [latex]\partial B^{\<i>}[/latex]-denoted by black and green curves representing the intersections of [latex]\partial \mathcal{E}(B) \cap \Sigma[/latex] and [latex]\partial \mathcal{E}(B^{\</i>}) \cap \Sigma[/latex], respectively-ensures the geometric integrity of these regions.](https://arxiv.org/html/2602.04888v1/x1.png)
New research constructs a powerful graph model to understand how entanglement dictates the limits of information accessible in evolving universes.
Shielding Federated Learning from Bad Actors
A new approach combines encryption and inference to defend against malicious participants in distributed machine learning.
Beyond the Bottleneck: Achieving Agreement with Less Communication
A new theoretical result demonstrates that reaching consensus doesn’t necessarily require the same communication overhead as broadcasting that consensus to all parties.
Clockwork Universe: Optimizing Space-Based Networks for Gravitational Wave Detection
A new analysis demonstrates how strategic geometric arrangements of optical lattice clocks can enhance the sensitivity of networks designed to detect the subtle ripples of the gravitational wave background.
The Echo of Errors: Predictable Vulnerabilities in AI-Generated Code
New research reveals that code created by artificial intelligence systems consistently repeats the same security flaws, creating opportunities for proactive attack prediction.
Decoding Patch Intent: AI Spots Bugs in Software Updates
Researchers have developed a new AI-powered technique that uses natural language documentation to automatically verify the correctness of software patches and identify potential vulnerabilities.
Frozen Waves: How Interactions Slow Down Material Evolution
![The post-quench Holstein model exhibits three distinct dynamical regimes-nonequilibrium metallic, quasi-coarsening, and arrested charge density wave (CDW) order-determined by the electron-phonon coupling strength λ, with transitions occurring near critical values of [latex]\lambda_{c1} \approx 0.4[/latex] and [latex]\lambda_{c2} \approx 1.0[/latex], where the system transitions from fluctuating CDW correlations to nucleation-limited coarsening and, ultimately, to dynamically arrested domain walls.](https://arxiv.org/html/2602.05815v1/x1.png)
New research reveals an unexpected slowdown in the coarsening process of charge-density waves, challenging conventional understanding of how materials evolve over time.
Beyond Trust: Building Resilient Ledgers for Imperfect Hardware
A new approach, Proteus, blends the speed of simpler consensus with the robust security of Byzantine Fault Tolerance to create ledgers that can withstand compromise in Trusted Execution Environments.
Learning Quantum Mechanics with Deep Neural Networks
Researchers have developed a fully differentiable framework to train machine learning models that can accurately predict both ground-state energies and excitation properties of molecules.