Existing black-box fingerprinting methods for large language models typically assume a single attack type, making them difficult to apply to realistic deployments involving multi-layer composite attacks. To address this limitation, this paper proposes a new black-box fingerprinting paradigm against composite attacks for commercial scenarios. For data construction, we design an LLM self-evaluation-based fingerprint probe selection mechanism. Candidate queries from multiple open-source datasets are fed into different base models, and a high-capability judge model is introduced to automatically assess output discrepancies across multiple dimensions, thereby selecting highly discriminative queries for constructing training and test sets. Methodologically, we reformulate black-box fingerprinting as a fine-grained identity authentication task. We employ a Transformer to model the semantic and logical features of question–answer sequences and incorporate ArcFace loss to impose angular-margin constraints. This compresses intra-class distributions of the same model under different attack conditions, enlarges inter-model decision boundaries, and learns more identity-invariant fingerprint representations. Furthermore, we systematically model and evaluate multi-layer composite attack chains composed of instruction tuning, quantization, distillation, RAG, and system prompts. Experimental results show that the proposed method significantly outperforms existing approaches under both single-layer and two-layer attack settings, and maintains over 61% Top-1 identification accuracy even in high-intensity three-layer composite attack scenarios.

Feature embedding has become a cornerstone technology for processing high-dimensional and complex data, which results in that Embedding as a Service (EaaS) models have been widely deployed in the cloud. To protect the intellectual property of EaaS models, existing methods apply digital watermarking to inject specific backdoor triggers into EaaS models by modifying training samples or network parameters. However, these methods inevitably produce detectable patterns through semantic analysis and exhibit susceptibility to geometric transformations including rotation, scaling, and translation (RST). To address this problem, we propose a novel fingerprinting framework called POSTER for EaaS models, rather than merely refining existing watermarking techniques. Different from watermarking, the proposed POSTER establishes EaaS model ownership through geometric analysis of embedding space's topological structure, rather than relying on the modified training samples or triggers. The key innovation lies in modeling the victim and suspicious embeddings as point clouds, allowing us to perform robust spatial alignment and similarity measurement, which inherently resists RST attacks. Experiments evaluated on visual and textual embedding tasks verify the superiority and applicability. This work reveals inherent characteristics of EaaS models and provides a promising solution for ownership verification of EaaS models under black-box scenarios.

Acceleration in deep reinforcement learning has attracted the attention of researchers. Many parallel training frameworks have been proposed to speed up the sampling and the training by running multiple agents simultaneously. However, the bottleneck of performance in a single agent is the neural network prediction which takes a lot of time as compared to the rapid environmental changes. Different from these parallel frameworks, we try to speed up the prediction of the agent to accelerate the entire training process. As far as we know, this is the first time to accelerate the training from this perspective. We propose a novel training framework NNC-DRL to accelerate the whole training process of deep reinforcement learning. NNC-DRL uses different neural networks to represent target policy and behavior policy. The behavior policy network is a smaller neural network derived from the target policy network, which could speed up the prediction. The target policy network will transfer its latest policy to the behavior policy network. The inconsistent policy distribution between behavior network and target network will degrade the convergence of the training. In NNC-DRL, we introduce the Important Sampling technique to estimate policy gradient, which could improve the convergence of the training. The experiments show that our approach NNC-DRL can speed up the whole training process by about 10-20% on Atari 2600 games with little performance loss.

In cloud datacenter, it should be rational to enforce fair allocation on network resources among VDCs (virtual datacenters) in terms of multi-tenant model. Traditionally, cloud networks are shared in a best-effort manner, making it hard to reason about how network resources are allocated. Prior works concentrate on either providing minimum bandwidth guarantee or achieving work-conserving based on the VM-to-VM flow policy or per-source policy, or both. However, fair allocation on redundant bandwidth among VDCs is ignored. In this paper, we design NXT-Freedom, a bandwidth guarantees enforcement framework that divides network capacity based on per-VDC fairness while achieving work-conservation. To ensure per-VDC fair allocation, a hierarchical max-min fairness algorithm is proposed. To be applicable to non-congestion-free network core and to be scalable, NXT-Freedom decouples computing per-VDC allocation from enforcing the allocation. Through evaluation of a prototype, we show that NXT-Freedom achieves per-VDC performance isolation, and can be rapidly adapted to flow variation in cloud datacenter.