Presentation Details
  1. Home
  2. Presentation List
  3. Presentation Details
HEBAP-FDS: A Hybrid Explainable Behavioral-Aware Privacy-Preserving Fraud Detection System
ID：58 View Protection：ATTENDEE Updated Time：2026-07-25 17:58:10 Hits：26 Online

Start Time：2026-07-30 14:55（Asia/Kolkata）

Duration：15min

Session：S3 Cyber Security » S3-2Cyber Security

Video No Permission Presentation File

Tips: Only the registered participant can access the file. Please sign in first.

Abstract
Financial fraud detection while operating in a federated context faces three principal challenges: The extreme imbalance of class distribution, The threat of adversarial evasion, Regulatory requirements about the privacy of data, including prohibitions against sharing of raw data. In providing a solution to these obstacles, the Hybrid Explainable Behavioural-Aware Privacy Preserving Fraud Detection System (HEBAP-FDS) identifies one significant weakness with conventional anomaly detection federations known as the “Robustness Paradox”, and empirically demonstrates that utilizing standard Byzantine Robust Aggregate functions (e.g. Trimmed Mean, FedProx) to determine federated weights is detrimental to detection of financial fraud as they discard valuable minority class gradients and categorize those values as anomalies. Consequently, the overall F1-score decreases to 1.7%. The HEBAP-FDS does not require the use of conventional (generic) forms of robustness or synthetic methods of oversampling to mitigate this issue rather, it utilizes a native focal loss function, as well as a per-user analysis and derivation of a composite user fraud detection score.The use of behavioural Mahalanobis distance drift layer and personalised federated learning with a decoupled classifier is an alternative to overcome this significant decline of the F1-score. The ULB Credit Card Fraud benchmark has a very low percentage of cards used for fraudulent activity, only 0.172%. By analysing the performance of the HEBAP-FDS method using 10 separate sets of random seed trials, the HEBAP-FDS method achieves the best performance in federated learning (FL) with an F1 score of 54.3% and 66.0% with HEBAP-FDS method, while obtaining only 1.7% and 22.4% with the FedAvg method, respectively, on the Independent Test Sets of the Extreme and IEEE-CIS datasets, respectively. In addition, the operation of the HEBAP-FDS method enables the preservation of the rigidity of localised data privacy for individual users, and provides a secure defence against adversarial attacks through the use of FGSM and also provides regulatory compliant explainability through the use of the SHAP write-up method.
Keywords
fraud detection, federated learning, adversarial robustness, Personalized FL, behavioral analytics, explainable AI, privacy-preserving ML.
Speaker
DHIVYA MOHAN
Student SRM Institute of Science and Technology (SRMIST)

Sirigiri Charitha
Student SRM Institute of Science and Technology *

Submission Author
DHIVYA MOHAN SRM Institute of Science and Technology (SRMIST)
Sirigiri Charitha SRM Institute of Science and Technology *
Dr. G. Sujatha SRM Institute of Science and Technology *
Submit Comment
Verify Code Change Another
All Comments
Important Date

  • Conference Date

    Jul 30

    2026

    to

    Aug 01

    2026

  • Jul 28 2026

    Draft paper submission deadline

  • Aug 03 2026

    Registration deadline

Sponsored By
The United Societies of Science
Organized By
Kongunadu College of Engineering and Technology
Supported By
IEEE Section
IEEE Madras Section
Contact Information
Login to view full contact details
Previous Conferences