Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10 times higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness.

The long-context capability of the Large Language Models (LLM) has made significant breakthroughs, but \textit{the maximum supported context length in length extrapolation} remains a critical bottleneck limiting their practical applications. The constraint of context length in LLMs arises from the self-attention mechanism, which cannot effectively and efficiently capture the semantic relationships within infinitely long contexts via the limited pre-trained positional information and attention scope. In this work, we propose \textbf{ReAttention}, a training-free approach enabling LLM based on the self-attention mechanism to support an infinite context with a finite attention scope under sufficient memory resources. ReAttention performs the position-agnostic top-$k$ attention before the ordinary position-aware self-attention, freeing LLMs from the length extrapolation issue. We validate the performance of ReAttention on the LongBench, L-Eval, and InfiniteBench and demonstrate that it is on par with traditional methods. Furthermore, we also apply ReAttention on mainstream LLMs, including LLaMA3.1-8B and Mistral-v0.3-7B, enabling them to support context lengths of at least 1M and even expanding the context length of LLaMA3.2-3B-chat by 128$\times$ to 4M without any further training in Needle-In-A-Haystack tests. We also improve the efficiency of ReAttention with Triton and achieve an efficient extrapolation without additional overhead. The code is available at \url{https://github.com/OpenMOSS/ReAttention}.

Speculative decoding accelerates Large Language Model (LLM) inference by employing a small speculative model (SSM) to generate multiple candidate tokens and verify them using the LLM in parallel. This technique has been widely integrated into LLM inference serving systems. However, inference requests typically exhibit uncertain execution time, which poses a significant challenge of efficiently scheduling requests in these systems. Existing work estimates execution time based solely on predicted output length, which could be inaccurate because execution time depends on both output length and token acceptance rate of verification by the LLM. In this paper, we propose a semi-clairvoyant request scheduling algorithm called Least-Attained/Perceived-Service for Speculative Decoding (LAPS-SD). Given a number of inference requests, LAPS-SD can effectively minimize average inference latency by adaptively scheduling requests according to their features during decoding. When token acceptance rate is dynamic and execution time is difficult to estimate, LAPS-SD maintains multiple priority queues and allows request execution preemption across different queues. Once the token acceptance rate becomes stable, LAPS-SD can accurately estimate the execution time and schedule requests accordingly. Extensive experiments show that LAPS-SD reduces inference latency by approximately 39% compared to state-of-the-art scheduling methods.

To protect sensitive data in cloud outsourcing smart systems, we implemented a private function evaluation system with fully homomorphic encryption (FHE) and trusted execution environment (TEE) using lookup tables (LUT), which preserves the privacy of data in various smart systems, including anomaly detection in smart grids. The proposed system introduces a trusted authority (TA) to manage FHE keys and generate private information retrieval (PIR) queries to extract the output from the LUT. The remaining problem in our previous work is that the TA knows the index distribution of data points in LUTs and the index of matched data points, which may cause security issues. In this work, we solve the security issues by executing operations in the TA with the TEE and reducing the runtime using unencrypted LUTs. Our system allows the computation of arbitrary multi-input functions to improve efficiency and expand the scope of applications with FHE in smart systems. We confirmed the evaluation runtime of an 18-bit one-input function needs 2.49 s. Compared to our previous work, our system achieved 2.1x speed up for an 8-bit three-input function with one thread.

In an advanced metering infrastructure (AMI), the electric utility collects power consumption data from smart meters to improve energy optimization and provides detailed information on power consumption to electric utility customers. However, AMI is vulnerable to data falsification attacks, which organized adversaries can launch. Such attacks can be detected by analyzing customers' fine-grained power consumption data; however, analyzing customers' private data violates the customers' privacy. Although homomorphic encryption-based schemes have been proposed to tackle the problem, the disadvantage is a long execution time. This paper proposes a new privacy-preserving data falsification detection scheme to shorten the execution time. We adopt elliptic curve cryptography (ECC) based on homomorphic encryption (HE) without revealing customer power consumption data. HE is a form of encryption that permits users to perform computations on the encrypted data without decryption. Through ECC, we can achieve light computation. Our experimental evaluation showed that our proposed scheme successfully achieved 18 times faster than the CKKS scheme, a common HE scheme.

In advanced metering infrastructure (AMI), the customers' power consumption data is considered private but needs to be revealed to data-driven attack detection frameworks. In this paper, we present a system for privacy-preserving anomaly-based data falsification attack detection over fully homomorphic encrypted (FHE) data, which enables computations required for the attack detection over encrypted individual customer smart meter's data. Specifically, we propose a homomorphic look-up table (LUT) based FHE approach that supports privacy preserving anomaly detection between the utility, customer, and multiple partied providing security services. In the LUTs, the data pairs of input and output values for each function required by the anomaly detection framework are stored to enable arbitrary arithmetic calculations over FHE. Furthermore, we adopt a private information retrieval (PIR) approach with FHE to enable approximate search with LUTs, which reduces the execution time of the attack detection service while protecting private information. Besides, we show that by adjusting the significant digits of inputs and outputs in our LUT, we can control the detection accuracy and execution time of the attack detection, even while using FHE. Our experiments confirmed that our proposed method is able to detect the injection of false power consumption in the range of 11–17 secs of execution time, depending on detection accuracy.

Smart grids are indispensable applications in smart connected communities (SCC). To construct privacy-preserving anomaly detection systems on a smart grid, we adopt fully homomorphic encryption (FHE) to protect users' sensitive data. Although FHE allows a third party to perform calculations on encrypted data without decryption, FHE only supports addition and multiplication on encrypted data. In anomaly detection, we must calculate both harmonic and arithmetic means consisting of logarithms. A naïve implementation of such arithmetic operations with FHE is a bitwise operation; thus, it requires huge computation time. To speed up such calculations, we propose an efficient protocol to evaluate any functions with FHE using a lookup table (LUT). Our protocol allows integer encoding, i.e., a set of integers is encrypted as a single ciphertext, rather than using bitwise encoding. Our experimental results in a multi-threaded environment show that the runtime of our protocol is approximately 51 s when the size of the LUT is 448,000. Our protocol is more practical than the previously proposed bitwise implementation.