AsiaJCIS 2026: Papers with Abstracts

Abstract. Most IoT malware operates as botnets dependent on Command and Control (C2) servers, but the short-lived nature of attack infrastructure often leaves samples dormant without C2 communication, hindering dynamic analysis. This paper proposes a system that combines Ghidra with a Large Language Model (LLM) to extract communication specifications from a malware binary and automatically generate a pseudo-C2 server. Experiments using Mirai demonstrate that the proposed system semantically interprets binary control structures and extracts all 20 core protocol elements in agreement with the ground truth (100% specification extraction accuracy). The generated pseudo-C2 server reproduces seven of ten DDoS attack vectors with full fidelity, generating traffic volumes of 82.9%—126.6% (averaging 100.1%) of the original C2. When applied to a customized variant created by modifying the publicly available Mirai source code, the method succeeds end-to-end — from specification extraction through pseudo-C2 generation to attack reproduction — demonstrating that the LLM infers specifications from binary structures without relying on pre-trained knowledge. This approach extends the applicability of LLMs from analysis assistance to the automated construction of dynamic analysis environments.

Abstract. Conventional Intrusion Detection Systems (IDSs) are often deployed in static environments, where the defender's view of attackers is largely limited to direct evidence of observed attack attempts, such as request patterns, payload characteristics, and the alerts they trigger. Although such information is useful for detection, it provides only a limited view of how an attack is carried out by an attacker against a target system. Moving Target Defense (MTD) can expand this view by repeatedly invalidating attacker knowledge and forcing adversaries to disclose additional behavioral reactions, such as renewed scanning, access to expired targets and decoy interaction. However, existing studies have mostly discussed such reactions at the level of detection or visibility enhancement, while their potential for further attacker profiling remains insufficiently explored.
This paper proposes an MTD-aware attacker profiling framework that transforms MTD-elicited reactions into structured behavioral evidence. By incorporating MTD service-state knowledge into IDS observation, the framework enables attacker interactions to be interpreted according to their mutation context rather than as isolated traffic events. We evaluate the framework in a controlled Docker-based web testbed using multiple simulated attackers with different behavioral tendencies. The results show that, compared with static IDS observation without MTD, mutation-aware IDS under MTD-driven mutation provides additional behavioral dimensions that make attacker differences more interpretable and distinguishable, enabling profiling-oriented analysis beyond visibility enhancement.

Abstract. In recent years, fake shopping sites have appeared at the top of search results through SEO poisoning, causing increasing damage. Conventional collection methods rely on fixed keywords and cannot keep up with evolving attack campaigns, delaying the discovery of new sites. We propose a closed-loop crawler that incorporates the page-level outputs of a fake-site classifier (fastText+LightGBM) into the search queries of the next cycle. Search queries are generated by a seed-compound strategy that combines characteristic words extracted from positive pages with seed words from the fake-shopping context (e.g., "deep discount", "official"). To complement evaluations that tend to focus on classifier accuracy, we also introduce per-cycle new-host counts and cumulative unique-host counts as exploration-range metrics. In a comparative experiment (n=3 for the proposed method, n=2 for the baseline), the fixed-keyword baseline yielded zero new-host acquisition from cycle 2 onward, indicating complete stagnation, whereas the proposed method continued to discover new hosts and, at cycle 3, achieved a cumulative unique-host count approximately 7.6 times that of the baseline on average.

Abstract. Phishing attacks targeting Japan have caused growing financial damage, motivating cluster-based analysis as a framework for prioritizing defensive actions. However, the longitudinal validity of such clusters — whether their definitions remain effective as phishing activity evolves — has not been empirically evaluated. We address this gap by applying cluster definitions and YARA rules established from 2023 data to 100,594 domains observed in 2024, and by comparing fixed 2023 rules against periodically updated 2024 operational rules. We find that 60.6% of 2024 domains are classified into the same clusters using only the 2023 rules, but trackability is non-uniform: high-impact clusters remain stable through Q1—Q2 and then degrade rapidly at the Q2—Q3 boundary, with the rule stable rate of high impact clusters dropping by an order of magnitude within a single quarter. The gap between fixed and updated rules splits into three operationally distinct components: rule degradation (18.0%), newly emerged clusters (18.3%), and classification conflicts (3.1%) — of which the first two dominate at comparable scale. Characterizing clusters along two axes — stability category (High/Medium/Low) and temporal pattern (Stable/Degrading/Volatile) — we provide quantitative guidance for scheduling rule maintenance.

Abstract. The rapid proliferation of malicious websites poses a significant threat to Internet users. While URL-based blocklists are widely used as a practical defense mechanism, their effectiveness depends not only on identifying newly emerging malicious URLs but also on maintaining the blocklist appropriately over time. This study evaluates the long-term operational effectiveness of a URL blocklist generated from real-user Web access logs.
From June 2024 to January 2026, we operated a URL blocklist generation method proposed in our previous work. We collected web access logs from real users and identified malicious URLs related to actual browsing activities. By adding these URLs and removing entries not accessed for more than two weeks, we maintained our blocklist every day and conducted a longitudinal cross-referencing experiment. Over 337 days, 6,866 malicious URLs were identified while maintaining an average blocklist size of 1,248 URLs. The blocklist matched 2,817 accesses from 102 users. We also found that 17% of the malicious URLs were removed soon after, whereas some URLs remained active for long periods. In addition, accesses to URLs after removal revealed limitations of simple lifetime-based maintenance. Finally, we conducted a real-user deployment in which the blocklist was updated daily and distributed weekly. Over 569 days, 13,208 malicious URLs were identified, while the average blocklist size was maintained at 1,330 URLs. During the 20-month evaluation, 84 users accessed URLs included in the blocklist 844 times. These results demonstrate that continuous removal of URLs can provide practical protection in real browsing environments while maintaining a relatively small blocklist size.

Abstract. We propose a ring signature scheme whose unforgeability is based on the assumption that the subfield bilinear collision (SBC) problem is hard. The SBC problem was introduced at CRYPTO 2024, which is related to a bilinear form over a finite field with a symmetric solution in its subfield. It is considered to be hard even against the power of quantum computers, and hence our ring signature scheme is expected to be a post-quantum one. The size of our ring signature is linear to the size of a ring, but shorter than a naive approach that uses the OR-proof technique. Concretely, when the size of a ring is 32, 128-bit quantum security is attained with the signature length being about 68 KB.

Abstract. Recently, Convolutional Neural Networks (CNNs) have demonstrated high performance in image recognition tasks, but they remain vulnerable to adversarial examples, which intentionally cause misclassification through imperceptible perturbations. Conventional defense methods using Autoencoders (AEs) rely solely on anomaly detection based on reconstruction error or image purification. These approaches are often insufficient against White-box attacks, as attackers can bypass either of the defenses, making robust protection difficult. In this study, we propose "DSVDD-AE," a two-stage defense method that integrates anomaly detection based on feature distance in the latent space using Deep SVDD (DSVDD) with image purification via AE. This two-stage approach enables the detection of attacks that cannot be effectively purified, providing robust defense even under White-box environments. In our evaluation, we constructed two types of models using Contractive AE (CAE) and Variational AE (VAE) and verified their defense performance against PGD attacks on the CIFAR-10 dataset. Under a White-box attack environment, the proposed method (DSVDD-VAE) achieved a defense success rate of 30.28%, marking an improvement of approximately 30 percentage points compared to the existing Defense-VAE.

Abstract. As a foundational technique for actor analysis in threat intelligence, we study authorship attribution based on stylistic features. Practical post analysis must address text brevity, formulaic expressions, and growing author counts, motivating a cross-cutting comparison of lightweight feature-based methods and pre-trained language models across performance, stability, and computational cost. Using Rakuten Ichiba reviews, we compare four methods — TF-IDF with logistic regression (TF-IDF+LR), BERT embeddings with logistic regression (BERT-Emb+LR), BERT fine-tuning (BERT-FT), and metric learning with k-nearest neighbors (Metric+kNN) — under unified settings, scaling the number of authors U up to 1000.

In our experiments, the lightweight TF-IDF+LR based on character n-grams achieves accuracy comparable to or higher than the BERT-based methods at U>=20, while requiring roughly 1/30 of BERT-FT's computation time at U=100 and 1/3 at U=1000. At U=1000 with n=186 reviews per author, TF-IDF+LR retains a Top-10 accuracy of 0.821, useful for candidate retrieval (screening). Applying Collective Attribution — a test-time aggregation in line with profile-based authorship attribution — to TF-IDF+LR, the concatenation variant improves accuracy at U=1000 from 0.623 with a single review (b=1) to 0.985 with ten reviews (b=10). For Japanese short web texts, our experiments suggest that combining lightweight character n-gram features with test-time aggregation offers a practical balance of accuracy, stability, and computational efficiency.

Abstract. In recent years, with the spread of teleworking and the expansion of public cloud computing, traditional perimeter-based security models that separate internal and external networks have become insufficient for adequate defense. Perimeter-based security models using traditional VPNs (Virtual Private Networks) and similar technologies have faced challenges: Once authenticated, they permit broad access to the internal network, and attacks on VPN devices directly expose the connected network to risk. Consequently, Zero Trust Network Access (ZTNA), based on the Zero Trust principle of not trusting any communication regardless of boundary presence, is gaining attention. ZTNA is recognized as a solution to the problems inherent in traditional network models by strictly controlling access for all communications, regardless of whether they originate inside or outside the organization, thereby minimizing user access.
However, ZTNA requires queries to an authorization server for every communication. In large-scale network environments, this leads to challenges such as increased processing load for access control and communication delays due to the rising number of queries. Furthermore, discussions on ZTNA have primarily focused on its concept and security model, with insufficient evaluation from the perspective of network performance.
This research constructs a ZTNA verification environment using the network emulator Mininet to evaluate network performance. Specifically, it adopts a structure using Envoy Proxy for the data plane and Open Policy Agent (OPA) for the authorization engine. As evaluation parameters, we measure the impact of the network distance (communication delay) between Envoy and OPA on the round-trip time (RTT) of packets during the authorization process, as well as changes in throughput and system load when increasing the number of concurrent authorization requests (session count).

Abstract. Secret sharing schemes are an important class of cryptographic protocols, and linear codes are one of the main tools for constructing such schemes. This paper presents a matroid characterization and the construction of an ideal secret sharing scheme for complete t-partite k-uniform hypergraph access structures. First, we present a representable connected matroid and prove that the complete t-partite k-uniform hypergraph access structure is exactly the port of this matroid. Then, based on the theory of Brickell and Davenport on matroids and ideal access structures, and using linear codes as a tool, we give an ideal secret sharing scheme that realizes this access structure. This scheme unifies Shamir's threshold scheme and the complete t-partite graph scheme of Brickell et al. as special cases.