18 August 2023 in IEEE Access
Since cyberattacks have become sophisticated in the form of advanced persistent threats (APTs), predicting and defending the APT attacks have drawn lots of attention. Although there have been related studies such as attack graphs, Hidden Markov Models, and Bayesian networks, they have four representative limitations; (i) non-standard attack modeling, (ii) lack of data-driven approaches, (iii) absence of real-world APT dataset, and (iv) high system dependability. In this paper, we propose Bayesian ATT&CK Network (BAN) which is based on system-independent data-driven approach. Specifically, BAN is based on Bayesian network, which adopts structure learning and parameter learning to model APT attackers with the MITRE ATT&CK® framework. The trained BAN aims to predict upcoming attack techniques and derives corresponding countermeasures. In addition, we prepare datasets via both automatic and manual labeling to overcome the data insufficiency issues of APT prediction. Experimental results show that BAN successfully contributes to handling APT attacks, given the best parameters extracted from extensive evaluations.
1 December 2021 in IEEE Communications Letters
Deep learning has been widely applied to automatic modulation classification (AMC), and there have been many studies on data augmentation techniques using deep generative models to improve performance. However, existing solutions need to train different models independently for each SNR, which leads to undeniable overhead. This letter presents UniQGAN, Unified Generative Adversarial Networks for IQ constellations of various SNRs, requiring a single model training. The proposed method introduces multi-conditions embedding and multi-domains classification to leverage both conditions, i.e., modulation type and SNR. Experimental results show that UniQGAN effectively improves the AMC performance, while the training time is reduced.
1 June 2023 in ICT Express
13 September 2023 in IEEE Access
The digital transformation of various systems has brought great convenience to our daily lives, but it has also increased the level of cyberattacks. As the number of cyberattacks has increased, so has the number of reports analyzing them, MITRE publishes the ATT&CK Matrix which analyzes the tactics and techniques of attacks based on real-world examples. As the flow of attacks has become more understandable through TTP information, researchers have been using it with deep learning models to detect or predict attacks, which makes embedding essential to train the model. In previous studies on embedding TTPs, embedding is limited to simple statistical methods such as one-hot encoding and TF-IDF. Such methods do not consider the order of TTPs and the conceptual similarity between TTPs, therefore do not capture the rich information that TTPs contain. In this paper, we propose embedding TTP with GloVe, a method using a co-occurrence matrix. To properly evaluate the semantic embedding performance of TTP, we also propose a measurement called Tactic Match Rate (TMR). In the experimental results, 8 out of 14 tactics showed a TMR of more than 0.5. Especially the “TA0007 (Discovery)” tactic showed the highest TMR of 0.87. Through correlation analysis, the experimental result shows that the reason for the different embedding performances of the tactic is affected by the frequency of the technique in the same tactic, with at most a 0.96 score. We also experimentally demonstrated that the neutrality of TTP affects learning performance.
27 March 2023 in IEEE Transactions on Dependable and Secure Computing
Automatic modulation classification (AMC) has been envisioned as a significant element for security issues at the physical layer due to its indispensable role in accurate communications. Recent attention to deep learning has impacted the AMC, which exhibits exceptional performance without manual feature engineering. To guarantee the accuracy and robustness of deep learning-based AMC, data augmentation is a critical issue. While existing studies have used several deep generative models to handle the data insufficiency, these studies face three challenges including low scalability, lengthy training time, and limited accuracy improvement. To this end, this paper presents UniQGAN, a novel unified generative architecture that models I/Q constellation diagrams from various signal-to-noise ratios (SNRs) using a single model. The proposed method enables the generation of high-quality data with a scalable generator, while requiring reduced training time. At the core of UniQGAN are multi-conditions embedding and multi-domains classification techniques that leverage both SNR and modulation type during the optimization process to enable unified modeling. Using abundant high-quality training data, UniQGAN accelerates the enhanced AMC with high performance and adversarial robustness. Experimental results demonstrate that the data generation by UniQGAN achieves superiority in terms of scalability, training time, and accuracy.
9 June 2025 in IEEE Transactions on Information Forensics and Security
As cyberattack operators have progressed to encompass group and nation-state levels, the nature of attacks has evolved into more sophisticated forms such as cyber campaigns. In response to these large-scale campaigns, tactical cyber threat intelligence (CTI) which focuses on tactics, techniques, and procedures (TTPs) has gained significant attention. However, the data-driven aspects of tactical CTI confront two primary challenges: (i) the extreme scarcity of campaign data and (ii) the difficulty of effectively integrating security domain knowledge. To this end, this paper presents MuCamp, a novel campaign generation method that operates in the context of limited campaign data while also considering the unique characteristics of large-scale attacks. The proposed method assumes that campaigns are TTP sequences, and based on this assumption, it generates valid campaign variants by replacing target TTP words with TTP synonyms, and preserves the strategic goals of the seed campaigns. MuCamp offers a scalable and interpretable augmentation strategy, enhancing CTI effectiveness under data scarcity and facilitating rapid adaptation to evolving threat landscapes. We also prepared a dataset consisting of 858 real-world campaigns labeled by security experts, including 14 tactics and 206 techniques, enabling reliable performance evaluation. Experimental results demonstrate that each component of MuCamp contributes to embedding-based group attribution by improving the separability of the correct group from alternative candidates, while effectively reflecting domain knowledge.