Existing hierarchical text classification (HTC) methods typically use prompt tuning or contrastive learning to inject the label hierarchy into a model as prior knowledge to implicitly learn label embeddings for classification.However, such implicit learning fails to accurately reflect label geometry (i.e., feature spatial distribution of label embeddings), as it does not model hierarchy-aware geometric relations among labels.To address this issue, we propose a novel two-stage label geometry structuring and aligning framework, termed LGSA, which transforms the label hierarchy from an implicit prior into an explicit embedding.First, we propose a hierarchical geometric structuring (HGS) module that leverages a general orthogonal frame (GOF) to reconstruct an explicit label geometry conforming to the label hierarchy.The label geometry is then treated as a label prototype to guide model training.To facilitate the guidance, we thereby propose a hierarchical geometric aligning (HGA) module as a regularization term to align label geometry learned by the model with the explicit label prototype.Experiments on three realworld HTC datasets confirm that LGSA consistently outperforms existing state-of-the-art methods.

Compressional SSDs can be a mixed blessing. While they offer users expanded logical space beyond the physical capacity, they complicate the Flash Translation Layer (FTL) design by requiring a larger Logical-to-Physical (L2P) address mapping, which places a heavier burden on the limited in-SSD memory, leading to a degraded I/O performance. In this paper, we aim to reduce the memory footprint of the L2P mapping table in compressional SSDs by proposing a novel N-to-1 L2P mapping table design that consolidates multiple logical entries into a single entry. This approach eliminates the duplication of physical page numbers when a physical page contains several compressed logical pages. To accommodate the dynamic compression ratios of real-world workloads, we introduce promotion and demotion that enable mapping table entries to migrate between pages with different compression ratios. Additionally, to address the issue of partial invalidation-where some compressed logical pages within a physical page are invalid due to the N-to-1 mapping-we present a compression-aware garbage collection algorithm aimed at minimizing the number of copy operations for partially invalid physical pages. We have implemented our design in MQsim, a widely used SSD simulator, and have conducted a series of experiments to evaluate the effectiveness of the proposed techniques. The results demonstrate that our approach significantly reduces the mapping table size in compressional SSDs, leading to an improved mapping table cache hit ratio and a reduced I/O latency compared to traditional compressional SSDs.

The rapidly growing mobile market is pushing flash storage manufacturers to expand capacity into the terabyte range. However, this presents a significant challenge for mobile storage management: more logical-to-physical page mappings are desired to be efficiently managed and cached while the available caching space is extremely limited. This motivates us to shift toward a new learning-based paradigm: rather than maintaining mappings for individual pages, the learning-based approach can represent mapping relationships for a set of continuous pages. However, to construct linear models, existing methods that either consume the already-limited memory space or reuse flash garbage collection demonstrate poor model construction capabilities or significantly degrade flash performance, making them impractical for real-world use. In this paper, we propose LFTL, a practical, learning-based on-demand flash translation layer design for flash management in mobile devices. In contrast to prior work that centered around gathering sufficient mappings for linear model construction, our key insight is that linear patterns can be extracted and refined by leveraging the orderly, LPA-aligned write stream typical of mobile devices. By doing this, highly accurate linear models can be constructed regardless of the constraints of mobile device's cache limitation. We have implemented a fully functional prototype of LFTL based on FEMU. Our evaluation results show that LFTL is more adaptable to memory-constrained storage devices than state-of-the-art learning-based approaches.

Machine learning as a service (MLaaS) has gained significant popularity and market traction in recent years, driven by advancements in Artificial Intelligence particularly Generative AI (GAI). However, MLaaS faces severe challenges from sampling-based model stealing attacks (MSAs), where attackers strategically query the targeted ML models provided by MLaaS providers to minimize the query burden while closely replicating the model’s functionality. Such MSAs pose severe consequences, including intellectual property (IP) theft and potential leakage of private training data. Unfortunately, existing defenses either sacrifice model utility or fail to generalize across diverse MSAs.In this paper, we propose DIARY, an innovative detection method specifically tailored to sampling-based MSAs by exploiting their inherent sophistication. Our key insight is that ‘clever’ malicious queries tend to extract more information from the targeted (victim) model than typical benign queries, as these attacks iteratively refine their queries by examining and analyzing prior queries and the corresponding responses. Hence we design DIARY to extract timing dependence within a query sequence and incorporate contrastive learning for properly characterizing such dependency that holds for different sampling-based MSAs. Comprehensive evaluations using five different sampling-based MSAs and two state-of-the-art defense baselines across four popular datasets consistently validate DIARY’s superior performance.

The Contrastive Language-Image Pretraining (CLIP) model has significantly advanced vision-language modeling by aligning image-text pairs from large-scale web data through self-supervised contrastive learning. Yet, its reliance on uncurated Internet-sourced data exposes it to data poisoning and backdoor risks. While existing studies primarily investigate image-based attacks, the text modality, which is equally central to CLIP's training, remains underexplored. In this work, we introduce ToxicTextCLIP, a framework for generating high-quality adversarial texts that target CLIP during the pre-training phase. The framework addresses two key challenges: semantic misalignment caused by background inconsistency with the target class, and the scarcity of background-consistent texts. To this end, ToxicTextCLIP iteratively applies: 1) a background-aware selector that prioritizes texts with background content aligned to the target class, and 2) a background-driven augmenter that generates semantically coherent and diverse poisoned samples. Extensive experiments on classification and retrieval tasks show that ToxicTextCLIP achieves up to 95.83\% poisoning success and 98.68% backdoor Hit@1, while bypassing RoCLIP, CleanCLIP and SafeCLIP defenses. The source code can be accessed via https://github.com/xinyaocse/ToxicTextCLIP/.

Modern SSD firmware is continuously optimized for higher parallelism to match the growing frontend PCIe bandwidth with more backend flash channels. Although a multi-core microprocessor is typically adopted to concurrently process independent NVMe requests from multiple NVMe queues, the existing one-to-many thread-request mapping model with each thread serving one or more incoming I/O requests has poor scalability due to severe lock contention problem, especially in cache management.

For enterprise storage systems, users' directory/file and IO access traces are critical for fine-tuning and new designs. However, once these systems are deployed, only trace features with small sizes are allowed to be sent back to vendors. Therefore, it is crucial to develop effective techniques for highly compressed feature extraction and feature-based high-fidelity trace regeneration. Existing works primarily focus on I/O trace modeling and regeneration without considering the directory/file access information. In this paper, we propose a new technique, called Sketcher, that can sketch massive traces into highly compressed “joint features” with both directory/file and I/O characteristics, and then based on these features regenerate high-fidelity traces with a learning-based approach. For trace feature extraction, one key idea is to divide traces into multiple distance-associated segments, where each segment contains all files and IO accesses operating under the same directory and the differences between segments are represented as displacement of segment inside the directory tree. A dynamic weight scaling technique is proposed to further compress features considering feature criticality and the size quota, thereby achieving high compression ratios with critical characteristics (e.g., abnormal IO access patterns). For trace regeneration, a new learning-based RNN model is proposed to regenerate high-fidelity traces from extracted features based on sampling directory trees. We have implemented a fully functional prototype based on typical enterprise storage systems and evaluated Sketcher with real applications and benchmarks on Huawei OceanStor Dorado storage server. Results show that Sketcher can effectively extract features with marginal runtime overheads while achieving compression ratios up to 15.2K and regenerating high-fidelity traces.

The continuously increasing volume of blockchain data presents significant challenges to blockchain traceability. Current tracing approaches, which rely on heuristic analytics directly applied to blockchain data, exhibit degraded tracking accuracy due to the involvement of only partial data. On the other hand, incorporating all blockchain data for analytics is not scalable given the unprecedented data volumes. Interestingly, we observe that despite the vast amount of blockchain data, blockchain tracing primarily focuses on time-related transaction spaces rather than all transactions. This insight motivates us to rethink the blockchain tracing problem to achieve both accuracy and scalability.

In recent years, Log-Structured Merge Tree (LSMtree) based key-value stores, such as LevelDB and RocksDB, have been widely adopted in data center systems. Though optimized for high-speed write processing, the severe I/O amplification remains a critical constraint that hinders them from reaching their maximum performance potential. Unfortunately, this problem is deeply rooted in the fundamental design of the LSMtree structure. A small number of frequently updated key-value items could quickly pollute the entire tree structure, causing repeated changes in the structure and quickly amplifying the amount of disk IOs across the levels in the tree. In this paper, we present a novel scheme, called Log-assisted LSM-tree (L2SM), to fundamentally address the long-existing I/O amplification problem. L2SM adopts a small-size, multi-level log structure to isolate selected key-value items that have a disruptive effect on the tree structure, accumulates and absorbs the repeated updates in a highly efficient manner, and removes obsolete and deleted key-value items at an early stage. We have prototyped the L2SM structure based on LevelDB. Our evaluation with the YCSB benchmark shows promising results by reducing the amount of disk IOs by up to 40.2%, increasing the throughput by up to 67.4%, and decreasing the average latency by up to 40.1%.