NTT DOCOMO

Transfer learning is crucial in training deep neural networks on new target tasks. Current transfer learning methods always assume at least one of (i) Source and target task label spaces overlap, (ii) Source datasets are available, and (iii) Target network architectures are consistent with source ones. However, holding these assumptions is difficult in practical settings because the target task rarely has the same labels as the source task, the source dataset access is restricted due to storage costs and privacy, and the target architecture is often specialized to each task. To transfer source knowledge without these assumptions, we propose a transfer learning method that uses deep generative models and is composed of the following two stages: pseudo pre-training (PP) and pseudo semi-supervised learning (P-SSL). PP trains a target architecture with an artificial dataset synthesized by using conditional source generative models. P-SSL applies SSL algorithms to labeled target data and unlabeled pseudo samples, which are generated by cascading the source classifier and generative models to condition them with target samples. Our experimental results indicate that our method can outperform the baselines of scratch training and knowledge distillation.

Increasing attention to digital identity and self-sovereign identity (SSI) is gaining momentum. SSI brings various benefits to natural persons, such as owning controls; conversely, digital identity systems in the real world require Sybil-resistance to comply with anti-money laundering (AML) and other needs. CanDID by Maram et al. proposed that decentralized digital identity systems may achieve Sybil-resistance and preserve privacy by utilizing multi-party computation (MPC), assuming a distributed committee of trusted nodes. Pass et al. proposed the formal abstraction of attested execution secure processors (AESPs) while equipping hardware-assisted security in mobile devices has become the norm. We first describe our proposal to utilize AESPs for building secure Sybil-resistant SSI systems, the architecture with a set of system protocols Π ^{G att}, which brings drastic flexibility and efficiency compared to existing systems. In addition, we propose a novel scheme that enables users (holders) to request verifiers to verify their credentials without AESPs, and it further achieves unlinkability among credentials created for public verification. Our scheme introduces a simplified format for computed claims and commitment-based anonymous identifiers. We also describe a technique to utilize zero-knowledge membership proofs, in particular, “One-Out-of-Many Proofs” Σ-protocol by Groth andKohlweiss, which can prove the existence of an expected credential without identifying it. Along with other techniques, such as utilizing the BBS+ signature scheme, we demonstrate how our scheme can achieve its goals with the extended anonymous and Sybil-resistant SSI system protocols Π ^{G att}. Entitling unlinkability among derived credentials in the anonymous Sybil-resistant SSI results in proper privacy preservation.

A lot of the devices in the Internet of Things are sensors responsible for capturing environmental information and relaying it to a system or network. Oftentimes, it is important to know the location of these sensors to better contextualize the information received from them. However, the sensors are purposely simple and cheap, meaning that conventional localization techniques, such as Global Navigation Satellite Systems are not feasible. Research has been done on using Unmanned Aerial Vehicles to estimate the location of the sensors, but issues with signal strength fluctuation and location approximation because of it are still prevalent. In this work, we propose a new method for estimating the location of sensors by exploiting the characteristics of radio wave signal propagation and creating a new flight path design, a solution to minimize the impact of variation in measurements, a novel candidate point generation through signal strength analysis, and a method to find the location based on the candidates. Through a thought-out experiment, we show that the proposed algorithm is overall significantly better than the existing solution when it comes to identifying the location of outdoor sensors.

In oral communication, especially public speaking, it is essential to speak at a speed that is appropriate for the situation. However, speech control requires substantial training. Although several speech-training systems that provide automated feedback on users' speech quality or behavior have been developed, users are still required to consciously control their ways of speaking to improve their speech. This study proposes a speech-supporting system that enables users to speak at a pace close to the target rate with minimum conscious adjustment. Because auditory feedback on the speaker's voice with a short delay disturbs speaking, we used auditory feedback with continuously varying delay times to slow the speaker's speech rate when speaking fast. We implemented a prototype and conducted a user study with ten speakers to confirm the effects on the speech style of the speaker during public speaking. The results show that the proposed system can slow the speech rate when the user speaks quickly without instructing the speaker on how to respond to auditory feedback. The findings also suggest that the proposed system causes less discomfort to the speaker than delayed auditory feedback with a constant delay.