Research axis: Week of December 2, 2024

Adaptive Security, Erasures, and Network Assumptions in Local Communications MPC

Multiparty n-party computation (MPC) is a cryptographic protocol technique that allows distinct parties to securely compute a function on their common data while keeping their inputs private. To build such a protocol, most work requires that all pairs of participating parties be able to communicate with each other securely and reliably. Recently, the MPC local communication (CL) problem has been explored in which this assumption is modeled more realistically – for example by only requiring that participating parties can communicate securely and reliably with a few other participating parties (as per example in networks like blockchains). . However, there are few solutions that ensure adaptive security (resilience to dynamic party corruption) and most rely on strong assumptions about party actions.

In a recent article: Adaptive Security, Erasures, and Network Assumptions in Local Communications MPCMicrosoft researchers and external collaborators revisit hypotheses made in previous work. The authors conclude that for a secure and adaptive CL-MPC, some previously assumed features (such as secure erase and multi-send) can be bypassed under certain conditions; However, completely reducing all-to-all communication to all remains unfeasible in CL contexts without some minimal assumptions. They propose a new SOS-RMT protocol, enabling more efficient CL-MPC within specific feasibility limits and additional cryptographic assumptions.

NEW RESEARCH

Seiche: a fair and predictable execution environment for cloud-hosted financial exchanges

Low-latency algorithmic trading improves the efficiency of modern financial markets by promoting accurate and timely pricing of securities, higher liquidity and lower trading costs for investors. The goal is to process incoming market data and issue trades as quickly as possible in order to take advantage of fleeting market-making and arbitrage opportunities. Interest in cloud-hosted financial exchanges is growing, as they promise, among other benefits, a cost-effective platform that is more accessible to market participants.

Unfortunately, one of the main hurdles in cloud environments is ensuring network and computation equality despite unpredictable network latencies as well as non-deterministic computation times.

In a recent preprint: Seiche: a fair and predictable execution environment for cloud-hosted financial exchangesMicrosoft researchers and external collaborators present a fair-by-design algorithmic trading platform that can operate in cloud environments. Cuttlefish aims to apply efficient and robust mapping of real operations to a new formulation of “virtual time”. This allows Cuttlefish to push fairness to the extreme, regardless of the underlying network communications and compute hardware. The researchers’ implementation and evaluation validate Cuttlefish’s practicality and demonstrate its operational effectiveness on public cloud platforms. This article builds on previous work: Rethinking financial exchanges hosted in the cloud for a fairer response time And DBO: Fairness for cloud-hosted financial exchanges.

NEW RESEARCH

LLM2CLIP: a powerful language model unlocks a richer visual representation

CLIP is a leading multimodal fundamental model, aligning visual and textual cues in a shared feature space. It supports various tasks including shotless classification, detection, segmentation and multimodal recovery, significantly influencing the entire multimodal domain. As a feature extractor, it has become dominant in multimodal representation tasks such as image understanding, video understanding, and text-image/video generation. However, rapid advances in large language models (LLMs) are continually pushing the boundaries of language understanding and generation. Can the capabilities of LLMs be exploited to further improve multimodal representation learning?

In a recent article: LLM2CLIP: a powerful language model unlocks richer visual representationMicrosoft researchers and external collaborators propose LLM2CLIP, a new approach to unlocking the potential of CLIP, focusing on fundamental optimizations of promising foundation models. By fine-tuning the LLM in the caption space with contrastive learning, they extract its textual capabilities in the output embeddings, significantly improving the textual discriminability of the output layer. The researchers then design a training process in which the refined LLM acts as a powerful teacher for CLIP’s visual encoder. The presence of the LLM allows them to incorporate longer and more complex captions without being limited by the pop-up window and capability limitations of the CLIP text encoder. Their experiments demonstrate that this approach provides substantial improvements to cross-modal tasks.

NEW RESEARCH

LORASC: Expressive and generalizable low-rank adaptation for large models via slow cascade learning

Core models, which are large-scale models pre-trained on large datasets and then adapted to specific tasks downstream, have become an integral part of contemporary machine learning frameworks. Fine-tuning these models is essential, but comprehensive parameter fine-tuning often faces significant memory and computational bottlenecks. Efficient parameter fine-tuning (PEFT) techniques aim to minimize the number of trainable parameters in order to reduce training costs and improve training stability. Among these techniques, low-rank adaptation (LoRA) is very effective, even if it has limits in terms of expressiveness and generalization.

In a recent article: LORASC: Expressive and generalizable low-rank adaptation for large models via slow cascade learningMicrosoft researchers and external collaborators present an innovative technique designed to improve the expressiveness and generalization capabilities of LoRA while preserving its training effectiveness. Their cascade learning strategy allows for a mixture of low-rank adaptations, thereby increasing the model’s ability to capture complex patterns. They also introduce a slow-fast update mechanism and noisy cascading tuning to strengthen generalization. Their extensive experiments on various language and vision datasets, as well as robustness tests, show that the proposed method significantly outperforms existing benchmarks, while mitigating overfitting, improving model stability and out-of-distribution robustness ( OOD).