💻 Computer Systems

Local energy communities—neighborhoods that share solar power among members—need accurate energy forecasting to balance supply and demand. Turazza et al. combine federated learning (privacy-preserving AI) with blockchain (transparent accounting) to enable peer-to-peer energy trading without exposing household consumption data.

The first malware families—PROMPTFLUX and PROMPTSTEAL—that invoke large language models at runtime have been documented, marking a shift from static attack scripts to adaptive, language-model-driven intrusion chains.

eBPF has become the foundation of Kubernetes autonomous security through tools like Cilium, Falco, and Tracee—yet the same kernel-level access that enables deep visibility also creates a concealment path for rootkits.

Every generation of AI hardware promises to solve the same three problems simultaneously: raw throughput, energy efficiency, and programmability. Every generation discovers that optimizing for two ...

Federated learning was supposed to solve the central tension of modern machine learning: you need large, diverse datasets to train good models, but the data you need is locked inside hospitals and ...

The U.S. Department of Defense has a problem measured in billions of lines of code.

Training a frontier large language model requires thousands of GPUs working in concert. The naive expectation is that doubling the GPUs should halve the training time.

Industrial control systems (ICS) and supervisory control and data acquisition (SCADA) networks manage the physical processes that underpin modern society—power generation, water treatment, oil and ...

Serverless computing simplifies deployment by abstracting infrastructure—but extending it to the edge introduces challenges in latency, scheduling, and resource heterogeneity. As LLM inference moves to edge devices, orchestrating serverless workloads across the cloud-edge continuum becomes a pressing systems challenge.

Zero-knowledge proofs allow one party to prove a statement is true without revealing any information beyond the statement's validity. Combined with blockchain, ZKPs are enabling privacy-preserving verification across domains—from academic credentials to financial KYC compliance and energy community governance.

Database administrators spend enormous effort tuning queries, indexes, and configurations. AI-driven autonomous database management systems aim to automate this entirely—using ML for predictive optimization, DRL for distributed query planning, and NLP for natural language database access.

Organizations want to train AI models on sensitive data in the cloud—but how do you trust the cloud provider? GPU Trusted Execution Environments create hardware-enforced enclaves where model weights and training data are encrypted even from the cloud operator. Lee et al. measure the performance cost.

Training large AI models on HPC clusters involves two under-exploited bottlenecks: the semantic coherence of training data and the interaction between distributed runtimes and heterogeneous hardware. SemanticHPC and DistZO2 propose solutions that go beyond standard data parallelism.

Formal methods—mathematically proving software correctness—have long been too expensive for general use. AI is changing the economics: LLM-assisted proof generation, automated test-case synthesis, and GenAI+formal methods synergy are making verification practical for automotive, aerospace, and security-critical software.

The maritime industry is undergoing rapid digitalization—autonomous vessels, IoT-connected cargo systems, satellite-dependent navigation. This digital transformation has exposed critical cybersecurity vulnerabilities that AI-driven threat detection is beginning to address.

Cloud-native applications built on microservices and containers present a different DDoS attack surface than traditional monolithic applications. Defending distributed systems requires rethinking mitigation at every layer—from ingress controllers to service mesh sidecars to inter-service rate limiting.

RISC-V's open instruction set architecture is gaining traction in AIoT devices—but existing operating systems are not optimized for real RISC-V hardware. Cheng et al. show how OS-level optimization can unlock the performance that RISC-V's flexibility promises for AI at the edge.

For decades, enterprises maintained separate databases for transactions (OLTP) and analytics (OLAP). HTAP systems promise to unify them—processing real-time transactions and complex analytics on the same data store. Kim et al. show how application-database co-design makes this practical.

Traditional centralized schedulers struggle at the cloud-edge boundary—latency to the control plane is too high, and single points of failure are unacceptable. Wen et al. propose using service mesh sidecar proxies as decentralized schedulers, turning infrastructure that already exists into intelligent orchestration agents.

Graph databases (Neo4j, TigerGraph, NebulaGraph) are growing rapidly—but their query optimizers harbor bugs that can silently produce incorrect results or catastrophic performance. Chen & Yu systematically analyze these bugs, revealing patterns that differ from those in traditional relational databases.

Global data center electricity consumption is growing rapidly, with AI workloads driving projections into the hundreds of terawatt-hours annually. Nunavath et al. propose an integrated framework combining sustainable workload scheduling, cloud-native efficiency, and edge-optimized inference to reduce computing's carbon footprint without sacrificing performance.

Connected vehicles generate massive volumes of network traffic that must be monitored for cyber intrusion—but pure neural network detectors are opaque and brittle. ZTID-IoV combines neurosymbolic AI (neural perception + logical reasoning) with federated meta-learning for adaptive, interpretable vehicle security.

Wearable AI devices monitor heart rate, sleep, activity, stress, and location continuously—generating intimate data streams that reveal health conditions, daily routines, and emotional states. Radanliev's framework addresses the urgent question: who is accountable when wearable AI systems misuse this data?

Cloud-native systems generate vast, heterogeneous security policies across containers, service meshes, API gateways, and serverless functions. Evaluating these policies for correctness and compliance is combinatorially explosive—and quantum optimization may provide the speedup needed for real-time evaluation.

Database query executors assume fixed memory allocations—but real workloads compete for memory dynamically. Otaki et al. propose resource-adaptive query execution that adjusts algorithms on the fly when memory pressure changes, preventing the catastrophic performance cliffs that occur when analytics queries spill to disk.

SQL remains the gatekeeping language of enterprise data—accessible to database specialists but opaque to the business users who most need data-driven insights. Multi-modal LLMs that translate natural language questions (and even dashboard screenshots) into database queries promise to democratize data access.

Synchronous request-response architectures are brittle—one slow service degrades the entire system. Event-driven architectures decouple services through message queues, absorbing traffic spikes and isolating failures. This methodology guide covers when to use EDA, how to design it, and what pitfalls to avoid.

A single whole-slide pathology image can exceed 10 gigapixels—far too large for any single GPU to process. ComPRePS 2.0 demonstrates how HPC clusters can process these images in parallel, enabling computational pathology at the scale needed for population-level cancer screening.

Social media generates millions of posts per minute. Extracting real-time sentiment from this firehose requires distributed NLP pipelines that parallelize text preprocessing, embedding, and classification across clusters—while maintaining sub-second latency for actionable insights.

Standard LLM fine-tuning requires storing model weights, gradients, optimizer states, and activations—often exceeding GPU memory for models above 70B parameters. DistZO2 eliminates backpropagation entirely, estimating gradients through forward-pass-only perturbation. Distributed across multiple GPUs, this enables fine-tuning of 100B+ models on hardware that cannot run standard training.

IoT devices generate commercially valuable data—traffic patterns, energy consumption, environmental conditions—but this data is trapped in silos because owners lack trusted mechanisms to share it. Consortium blockchains with ZKP enable data trading where buyers verify data quality without accessing the data itself.

Smart cities need bandwidth that wireless alone cannot provide. Adib et al. trace the full path from optical network architecture through physical fiber deployment, showing how fiber-to-the-premises enables the sensor density, data throughput, and latency requirements that define urban intelligence.

Financial transactions on public blockchains are transparent by design—but transparency conflicts with financial privacy. ZKP-enabled payment systems allow users to prove transaction validity and regulatory compliance without revealing amounts, counterparties, or account balances.

Academic credential fraud imposes significant costs on employers and undermines legitimate graduates. ZKBAR-V enables blockchain-anchored degree verification where employers confirm credentials without accessing any personal academic data—eliminating both fraud and privacy risks.

Health records must be shared across providers for care coordination—but sharing exposes sensitive patient data to breaches and misuse. Three 2025 systems demonstrate blockchain+ZKP architectures where patients control access, providers verify clinical data, and no centralized database stores the complete record.

Aging bridges are monitored by sensor networks, but raw sensor data reveals symptoms—not diagnoses. Digital twins that mirror bridge behavior in real time, continuously calibrated by genetic algorithms, can predict structural failures before they become visible, transforming infrastructure maintenance from reactive to predictive.

As AI-generated content becomes indistinguishable from human-created content, proving that an online entity is a unique human—not a bot or a duplicate—becomes a foundational infrastructure problem. Proof of personhood on blockchain offers cryptographic verification of humanity without revealing identity.

Confidential computing extends hardware-based trusted execution environments to GPU memory, enabling organizations to run AI training and inference on untrusted infrastructure without exposing models or data. With 75% of organizations reportedly adopting by 2025, the technology addresses a growing tension between cloud AI economics and data sovereignty.

Autoregressive decoding—generating one token at a time—remains the primary throughput bottleneck in LLM serving. Berkeley's integration of P-EAGLE parallel speculative decoding into vLLM generates K draft tokens in a single forward pass, with Eagle3 representing current state-of-the-art and TurboSpec adding closed-loop dynamic parameter control.

Running large language models at the edge—on devices rather than in data centers—can reduce inference energy consumption by up to 75% and costs by over 80%. This review examines the quantization techniques, model choices, and hybrid architectures that make on-device LLM inference practical.

Google reports that memory safety vulnerabilities have dropped below 20% of total Android vulnerabilities for the first time, driven by Rust adoption in new code. MIT, NSA, CISA, and DARPA's TRACTOR program signal that memory-safe languages are moving from recommendation to institutional mandate.

The discovery of Shai-Hulud 2.0—the first self-propagating worm in the npm ecosystem, infecting 500+ package versions—marks a shift from targeted supply chain attacks to autonomous propagation. OWASP's 2025 ranking of supply chain failures at #3 and the rise of SBOM as a build-tool primitive reflect the industry's belated reckoning with dependency trust.

Google's Willow processor achieved below-threshold quantum error correction with a 101-qubit distance-7 surface code, suppressing logical errors by a factor of 2.14 per code distance increment. We examine what this milestone means—and does not mean—for practical quantum computing.

Combining fully homomorphic encryption with federated learning promises ML training where no party—not even the aggregation server—can see raw gradients. The Lancelot framework demonstrates a 20x speedup over prior FHE-based FL methods while resisting Byzantine attacks, but significant overhead remains.

A new hybrid consensus architecture combines RAFT's throughput with PBFT's Byzantine fault tolerance in a two-stage pipeline, achieving near-RAFT performance under normal conditions while tolerating malicious nodes. Formal verification confirms safety and liveness properties.