Fully homomorphic encryption (FHE) is a powerful cryptographic primitive to secure outsourced computations against an untrusted third-party provider. With the growing demand for AI and the usefulness of machine learning as a service (MLaaS), the need for secure training and inference of artificial neural networks is rising. However, the computational complexity of existing FHE schemes has been a strong deterrent to this. Prior works suffered from accuracy degradation, lack of scalability, and ciphertext expansion issues. In this paper, we take the first step towards the problem of space-efficiency in evaluating deep neural networks through designing DOReN: a low depth, batched neuron that can simultaneously evaluate multiple quantized ReLU-activated neurons on encrypted data without approximations. Our circuit design reduced the complexity of the accumulator circuit depth from O(logm ·logn) to O(logm + logn) for n bit integers. The experimental results show that the amortized processing time of our homomorphic neuron is approximately 1.26 seconds for 300 inputs and less than 0.13 seconds for 10 inputs at 80 bit security, which is a 20 fold improvement upon Lou and Jiang, NeurIPS 2019.
Decentralized identity (DID) systems aim to give users full control over their identities by using completely decentralized technologies, such as blockchain or distributed ledgers, as identity providers. However, when user credentials are compromised, it is impossible in existing DID systems for the users to detect credential misuse. In this paper, we propose new DID authentication protocols with two properties: auditability and privacy. The former enables the detection of malicious authentication events, while the latter prevents an adversary from linking an authentication event to the corresponding user and service provider. We present two protocols that achieve auditability with varying privacy and performance guarantees. The first protocol has high performance, but it reveals information about the user. The second protocol achieves full privacy, but it incurs a higher performance overhead. We present a formal security analysis of our privacy-preserving protocols by using the Tamarin prover. We implemented them and evaluated their performance with a permissioned blockchain deployed over the Amazon AWS and a local cloud infrastructure. The results demonstrate that the first protocol is able to support realistic authentication workloads, while the second is nearly practical.
1 April 2022 in Future Generation Computer Systems
The appealing properties of secure hardware solutions such as trusted execution environment (TEE) including low computational overhead, confidentiality guarantee, and reduced attack surface have prompted considerable interest in adopting them for secure stream processing applications. In this paper, we revisit the design of parallel stream join algorithms on multicore processors with TEEs. In particular, we conduct a series of profiling experiments to investigate the impact of alternative design choices to parallelize stream joins on TEE including: (1) execution approaches, (2) partitioning schemes, and (3) distributed scheduling strategies. From the profiling study, we observe three major high-performance impediments: (a) the computational overhead introduced with cryptographic primitives associated with page swapping operations, (b) the restrictive Enclave Page Cache (EPC) size that limits the supported amount of in-memory processing, and (c) the lack of vertical scalability to support the increasing workload often required for near real-time applications. Addressing these issues allowed us to design SecJoin, a more efficient parallel stream join algorithm that exploits modern scale-out architectures with TEEs rendering no trade-offs on security whilst optimizing performance. We present our model-driven parameterization of SecJoin and share our experimental results which have shown up to 4-folds of improvements in terms of throughput and latency.
Agentic systems operating in adaptive automation environments require persistent memory to retain past actions, qualify outcomes, and improve performance over time. Existing memory mechanisms are architecturally diverse but evaluated in isolation using inconsistent benchmarks, limiting direct comparison. This scoping review systematically examines twelve state-of-the-art memory systems across four architectural categories: Structured Memory, Long-Term Memory, Short and Long-Term Memory, and Workflow and Procedural Memory. Eleven of the twelve are evaluated empirically across nine benchmark datasets spanning conversational question answering, long-document comprehension, and agentic task execution. Memory3 is surveyed but not evaluated, because its memory is built into the model through pre-training and cannot be run on a shared frozen backbone. Every system is evaluated on the datasets aligned to its architectural category, using the same GPT-4o-mini backbone, the same metrics, and the same decoding settings. We also run the backbone with no memory at all, which gives a floor against which each system’s contribution can be measured. Results show that no single architecture dominates all task types. Mem1 leads Contains-Match on conversational tasks while MemoryOS leads Token F1 on the same group, ReadAgent leads multiple-choice long-document comprehension and A-Mem leads open-ended long-document generation, and AWM beats AgentFly on all workflow-level tasks. Measured against the no-memory floor, memory helps a great deal on long-document benchmarks but adds little on multi-hop factoid retrieval, where several systems actually score below the backbone answering from its own pre-trained knowledge. We also identify a verbosity confound. Contains-Match rewards systems that write long answers, with a Spearman correlation of between response length and the gap between Contains-Match and Token F1, and it penalizes short exact answers when the reference string carries a trailing period or article. We recommend reporting Token F1 alongside Contains-Match in future evaluations. All evaluation code and datasets are publicly available; see the Data Availability Statement.