Kaoru Otsuka

Kaoru Otsuka

PhD student at OIST

Publications

You can also find my articles on my Google Scholar profile.

Conference Papers

Delayed Momentum Aggregation: Communication-efficient Byzantine-robust Federated Learning with Partial Participation

Published in Appeared at ICML 2026, 2026

Partial participation is essential for communication-efficient federated learning at scale, yet existing Byzantine-robust methods typically assume full client participation. In the partial participation setting, a majority of the sampled clients may be Byzantine, once Byzantine clients dominate, existing methods break down immediately. We introduce delayed momentum aggregation, a principle where the central server aggregates cached momentum from non-sampled clients along with fresh momentum from sampled clients. This principle ensures Byzantine clients remain a minority from the server’s perspective even when they dominate the sampled set. We instantiate this principle in our optimizer DeMoA. We analyze the convergence rate of DeMoA, showing that DeMoA is Byzantine-robust under partial participation. Experiments show that, with 20% Byzantine ratio and only 10% partial participation rate, DeMoA achieves the best accuracy even when existing methods fail empirically.

Download Paper

Shepherd - High-Precision Coverage Inference for Response-guided Blackbox Fuzzing (Registered Report)

Published in ISSTA Companion 2025 - 34th ACM SIGSOFT International Symposium on Software Testing and Analysis, 2025

In recent years, fuzzing has gained attention as a primary means for the early detection of vulnerabilities. Although coverage-based greybox fuzzing utilizes internal coverage information to achieve high exploration efficiency, it remains difficult to employ the fuzzing framework in some restricted environments where we cannot instrument the program, such as firmware or smartphone applications. In contrast, blackbox fuzzing does not require runtime information and is thus more widely applicable, but suffers from lower efficiency because coverage cannot be measured. To address this issue, there is a growing demand for methods that can approximate coverage in blackbox environments to optimize fuzzing. One existing study proposes estimating coverage based on the relationship between program responses and strings embedded in its binary. However, this approach faces challenges with ambiguous matching algorithms and the non-uniqueness that occurs when a single string is shared by multiple basic blocks, leading to frequent misestimations. In this research, we propose a new coverage inference method, Shepherd, which combines high-precision string matching with context analysis to resolve these problems. Experimental results show that Shepherd significantly improves estimation accuracy compared to the existing approach.

Download Paper