Ensuring the secure and proper deployment of Enclaves is critical for the security of Trusted Execution Environments (TEEs). Recently, the increasing attacks targeting Enclaves have made the traditional solution risky to use different enclaves for each task. An ideal solution is a pre-designed enclave that can be flexibly reused across different computational tasks. To address this issue, we propose FH-TEE, a novel framework that integrates Fully Homomorphic Encryption (FHE) with TEE. In this approach, computational tasks are partitioned so that only fixed, logic-independent operations are performed within the Enclave, enabling its reuse across various tasks. Flexibility in accommodating diverse computational requirements is achieved through FHE. Furthermore, because computationally intensive FHE operations, such as bootstrapping and ciphertext transformation, are logic-independent and offloaded to the Enclave, FH-TEE significantly improves efficiency compared to purely FHE-based solutions, while maintaining robust privacy and security. In our demo program, the server receives ciphertext from the client, performs ten iterations of the Game of Life computation, and then sends the resulting ciphertext back to the client. The results show that our scheme is 10.57× faster in computation speed and 32064× smaller in transmission size compared to a pure FHE scheme.
As artificial intelligence and deep learning models become increasingly sophisticated, processing vast amounts of personal data, the need for robust privacy protection measures has become paramount. This paper addresses the critical challenge of implementing privacy-preserving neural networks (PPNN) using fully homomorphic encryption (FHE), with a focus on the TFHE scheme. We present novel approaches to homomorphically implement maxpooling, a crucial non-linear operation in convolutional neural networks, which has been challenging to realize within existing FHE frameworks. Our main contributions include two innovative homomorphic maxpooling algorithms: HSMaxPool, utilizing the non-linear Sign function, and HRMaxPool, based on a homomorphic ReLU function. Both algorithms are implemented using TFHE’s gate bootstrapping and programmable bootstrapping techniques, respectively. Experiments were conducted on neural networks operating over ciphertext, employing the proposed homomorphic maxpooling algorithms. The results show that our methods deliver improved inference accuracy and reduced execution time compared to prior approaches, while maintaining accuracy throughout the maxpooling phase. This demonstrates the effectiveness of our homomorphic maxpooling algorithms in enhancing performance without sacrificing precision, further advancing the feasibility of efficient and privacy-preserving neural network computations on encrypted data.
A ring signature can be performed by anyone in a group of users who holds a signing key, and the signature of a message reveals nothing about the signer’s identity except the fact that they are a ring member. Currently, the size of most ring signatures depends on the number of ring members, but for realistic applications with a large number of ring members, this will lead to higher communication overheads. To reduce the signature size, in this paper, a new transparent compact ring signature scheme is constructed. We use homomorphic encryption and constrained signatures to improve the efficiency of the signature scheme when the ring size is large, thereby achieving “fully compact” signatures whose size is O(1), i.e., constant in the number of ring members. Not only does the scheme guarantee the privacy of the signer’s identity with less communication overhead, but it also provides stronger security under the RO model, including unforgeability and anonymity, under the assumption of the quantum hardness of the standard SIS problem. Furthermore, to the best of our knowledge, our scheme is the first compact post-quantum ring signature scheme that has transparent setup, requiring only public randomness (and not any trusted entity).
Identifying related Indicators of Compromise (IoCs), such as malicious IP addresses and associated network services (IP:Ports), is crucial for threat intelligence but challenging due to the scale and heterogeneity of Internet scan data. We propose a novel self-supervised contrastive learning framework to generate meaningful embeddings for IP:Port instances by fusing multimodal features (TLS certificates, banners, web content, metadata). Addressing the critical false negative problem inherent in contrastive learning, especially when combined with Hard Negative Mining (HNM), we introduce Semantic-Consistency HNM (SC-HNM). This technique employs a lightweight, integrated auxiliary network operating on early-stage feature representations to assess semantic similarity between an anchor and potential hard negatives. If the initial similarity exceeds a threshold δ , the potentially false negative is dynamically filtered out before computing the main contrastive loss, thus preserving semantic structure without requiring any labels. Our multimodal encoder utilizes attention and gating mechanisms for effective feature fusion. We train the model entirely from scratch on 8 million unlabeled scan records. Experiments show our approach yields highly competitive IoC retrieval performance, significantly outperforming zero-shot baselines (including large pre-trained models like BERT, RoBERTa, BGE-M3, E5). Ablation studies confirm the benefits of multimodality, HNM, and our SC-HNM filter. Furthermore, the resulting model is compact ( 27 million parameters) and efficient, enabling practical deployment for large-scale IoC similarity search and threat discovery.
