Deep neural networks (DNNs) have made outstanding achievements in a wide variety of domains. For deep learning tasks, large enough datasets are required for training efficient DNN models. However, big datasets are not always available, and they are costly to build. Therefore, balanced solutions for DNN model efficiency and training data size have caught the attention of researchers recently. Transfer learning techniques are the most common for this. In transfer learning, a DNN model is pre-trained on a large enough dataset and then applied to a new task with modest data. This fine-tuning process yields another challenge, named catastrophic forgetting. However, it can be reduced using a reasonable strategy for data argumentation in incremental learning. In this paper, we propose an efficient solution for the random selection of samples from the old task to be incrementally stored for learning a sequence of new tasks. In addition, a loss combination strategy is also proposed for optimizing incremental learning. The proposed solutions are evaluated on standard datasets with two scenarios of incremental fine-tuning: (1) New Class (NC) dataset; (2) New Class and new Instance (NCI) dataset. The experimental results show that our proposed solution achieves outstanding results compared with other SOTA rehearsal methods, as well as traditional fine-tuning solutions, ranging from 1% to 16% in recognition accuracy.

Phishing attacks by malicious URL/web links are common nowadays. The user data, such as login credentials and credit card numbers can be stolen by their careless clicking on these links. Moreover, this can lead to installation of malware on the target systems to freeze their activities, perform ransomware attack or reveal sensitive information. Recently, GAN-based models have been attractive for anti-phishing URLs. The general motivation is using Generator network (G) to generate fake URL strings and Discriminator network (D) to distinguish the real and the fake URL samples. This is operated in adversarial way between G and D so that the synthesized URL samples by G become more and more similar to the real ones. From the perspective of cybersecurity defense, GAN-based motivation can be exploited for D as a phishing URL detector or classifier. This means after training GAN on both malign and benign URL strings, a strong classifier/detector D can be achieved. From the perspective of cyberattack, the attackers would like to to create fake URLs that are as close to the real ones as possible to perform phishing attacks. This makes them easier to fool users and detectors. In the related proposals, GAN-based models are mainly exploited for anti-phishing URLs. There have been no evaluations specific for GAN-generated fake URLs. The attacker can make use of these URL strings for phishing attacks. In this work, we propose to use TLD (Top-level Domain) and SSIM (Structural Similarity Index Score) scores for evaluation the GAN-synthesized URL strings in terms of the structural similariy with the real ones. The more similar in the structure of the GAN-generated URLs are to the real ones, the more likely they are to fool the classifiers. Different GAN models from basic GAN to others GAN extensions of DCGAN, WGAN, SEQGAN are explored in this work. We show from the intensive experiments that D classifier of basic GAN and DCGAN surpasses other GAN models of WGAN and SegGAN. The effectiveness of the fake URL patterns generated from SeqGAN is the best compared to other GAN models in both structural similarity and the ability in deceiving the phishing URL classifiers of LSTM (Long Short Term Memory) and RF (Random Forest).

In recent years, text-based person search (TBPS) has garnered significant attention, which involves tackling complex issues in both natural language processing and computer vision. Although significant progress has been made, most TBPS methods have focused mainly on English-language descriptions. As a result, person search performance in non-English languages remains significantly lower compared to English. Multilingual models, which can leverage high-resource languages to enhance understanding in low-resource ones, offer a promising solution for TBPS. In this paper, we propose a multilingual text-based person search method, named TBPS-mSigLIP, which utilizes a multilingual visual-language model as an encoder to extract both visual and textual features. Then, we incorporate various data augmentation techniques and loss functions to enhance model performance. To assess the effectiveness of multilingual models across both high- and low-resource languages, we have conducted extensive experiments on four benchmark datasets including CUHK-PEDES, 3000VnPersonSearch, 3000Vn-V2E, and PRW-TPS-CN with queries in three languages (i.e. Vietnamese, English, and Chinese). The proposed method achieves Rank@1 accuracies of 70.76%, 49.70%, 39.33%, and 46.78% on CUHK-PEDES, 3000VnPersonSearch, 3000Vn-V2E, and PRW-TPS-CN datasets, respectively. It outperforms existing methods on both the 3000VnPersonSearch and 3000Vn-V2E datasets, with Rank@1 improvements of 9.4% and 11.05% over CFFVL [1] and M-IRRA [2] on 3000VnPersonSearch, and 2.26% and 3.93% over TBPS-CLIP and M-IRRA on 3000Vn-V2E, respectively. On CUHK-PEDES and PRW-TPS-CN, the proposed framework demonstrates competitive performance compared to state-of-the-art methods. Furthermore, we introduce a two-stage transfer learning paradigm to reduce memory usage and training time while maintaining search performance. Our code is available at: https://github.com/sigm-seee/TBPS_mSigLIP.

In recent years, many efforts have been dedicated to text-based person search, thanks to its potential applications in various domains. However, most of these works focus on person search via queries in English and conduct offline evaluations. Despite some promising results for text-based person search in English, several challenges still prevent its widespread use in practical situations when deployed in minor languages. This article extends person search to the Vietnamese language. In terms of linguistics, English and Vietnamese belong to two different language families. In addition to the difference in vocabulary, these two languages also have opposite word structures and syntactic structures. The contributions of the article are twofold. First, based on the network architecture of the ViTAA model [Wang et al. 2020 ], a framework for person search through Vietnamese queries has been developed. In this framework, to take into account specific characteristics of the Vietnamese language, the word-tokenizing, Parts of Speech (PoS) tagging techniques of different natural language processing tools, including Underthesea [UndertheseaNLP 2018 ], SEACoreNLP [Singapore 2021 ], and PhoNLP [Nguyen and Nguyen 2021 ], have been investigated in order to extract language elements from Vietnamese descriptions. Our investigation shows that selecting a suitable preprocessing technique can improve person search performance by 1.28% at R@1. When incorporating these preprocessing techniques with the person search model, the best accuracy was achieved with 27.08%, 51.38%, and 63.00% at rank 1, rank 5, and rank 10 respectively. Second, for the first time, an online evaluation of person search through natural language queries was conducted. A web-based application has been developed to serve online evaluation scenarios with different groups of end-users. An extensive evaluation was conducted with 30 subjects and 115 queries. Upon analyzing the experimental results, open issues and suggestions for future improvements in person search were uncovered.