Ana gezinime geç Aramaya geç Ana içeriğe geç

Deep Multi-Modal Learning for Automated Categorization and Evidence Triaging in Large-Scale Drone Forensic Archives

  • Azamat Baibussinov
  • , Aliya Ospanova
  • , Gulnara Abitova
  • , Kaisarbek Yesbergenov
  • , Ibraheem Shayea
  • , Aina Tazhiyeva*
  • *Bu çalışma için yazışmadan sorumlu yazar
  • National Defense University of The Republic of Kazakhstan
  • Republican State Enterprise On The Right of Economic Management Republic of Kazakhstan
  • Astana IT University

Araştırma çıktısı: Kitap/Rapor/Konferans Bildirisinde Bölüm › Konferans katkısı › Hakemli

Özet

The increasing volume and complexity of data stored in civilian Unmanned Aerial Vehicles (UAVs) necessitate a transition from manual forensic inspection to automated analytical frameworks. This paper introduces and evaluates a multi-modal deep learning framework for automated classification and triaging of forensic artifacts in large-scale drone memory archives. The framework employs a dual-stream architecture: a fine-tuned EfficientNet-B3 backbone for visual content analysis and a bidirectional LSTM network with self-attention for telemetry log sequence classification. An attention-based feature fusion module integrates both modalities into a unified evidence representation. The system was trained and evaluated on a curated dataset of 18 memory dumps (431.7 GB) from four UAV platforms (DJI Mavic 2 Pro, Phantom 4, Spark, and Parrot Bebop 2), acquired using the ISP bit-stream extraction methodology detailed in our companion work on hardware-level UAV forensics. The resulting dataset comprises 14,826 annotated samples across four forensic artifact categories. Using a leave-one-device-out cross-validation protocol, the framework achieved a weighted F1-score of 0.936 (95% CI: 0.921-0.951), with per-class F1-scores ranging from 0.908 to 0.967. An ablation study demonstrates that the attention-based fusion outperforms naive concatenation by 3.2 percentage points and that the multi-modal approach outperforms single-stream baselines by 5.8-11.4 pp. Gradient-weighted Class Activation Mapping (Grad-CAM) and LSTM attention weight visualizations provide interpretable evidence justifications. On a dual-GPU workstation, the framework processes a 32 GB memory dump in 1.8 hours, representing an 87% reduction in initial triaging time compared to manual expert analysis. The framework is designed as a decision-support tool for forensic examiners, not a replacement for expert judgment.

Orijinal dilİngilizce
Ana bilgisayar yayını başlığı2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications, ICCSDFAI 2026
YayınlayanInstitute of Electrical and Electronics Engineers Inc.
Sayfalar874-881
Sayfa sayısı8
ISBN (Elektronik)9798319520487
DOI'lar
Yayın durumuYayınlandı - 2026
Etkinlik2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications, ICCSDFAI 2026 - Istanbul, Türkiye
Süre: 25 Haz 2026 → 27 Haz 2026

Yayın serisi

Adı2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications, ICCSDFAI 2026

???event.eventtypes.event.conference???

???event.eventtypes.event.conference???2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications, ICCSDFAI 2026
Ülke/BölgeTürkiye
ŞehirIstanbul
Periyot25/06/26 → 27/06/26

Bibliyografik not

Publisher Copyright:
© 2026 IEEE.

Parmak izi

Deep Multi-Modal Learning for Automated Categorization and Evidence Triaging in Large-Scale Drone Forensic Archives' araştırma başlıklarına git. Birlikte benzersiz bir parmak izi oluştururlar.

Alıntı Yap