PRIVACY-PRESERVING PERSONALIZED TRAVEL PLANNING USING PHASE-ADAPTIVE FEDERATED LEARNING

Abstract

Background

Personalized travel itinerary systems are crucial for modern travelers, since they provide individualized suggestions based on choices like location, budgetary constraints, and interest.

Methodology

The phase-adaptive federated learning methodology takes a three-tiered approach to privacy-preserving journey planning, defining training phases (cold-start, warm-up, and convergence) using gradient norms and loss stability analysis. Client selection begins randomly and then moves to the top 40% of high update clients during convergence. Aggregation happens every 5 rounds rather than every round, and local epochs grow from 1 to 5. It safeguards sensitive user preferences by federating global parameters while retaining personalized layers on-device and employing phase-aware noise injections to thwart model inversion assaults.

Results

PAFL method produces three major findings: (1) Superior personalization with0.87nDCG@10, 18.7% higher than FedAvg and near centralized performance (0.89); (2) Strong privacy preservation with a <3.1% attack success rate in model inversion attacks, showing a 12x lower vulnerability than FedAvg (45.3%); and (3) Optimal efficiency with 42% communication reduction (59.4 MB total) through phase-adaptive aggregation. These findings provide new benchmarks for the privacy-utility trade-off in federated travel planning.

Conclusion

PAFL offers near-centralized recommendations quality while significantly reducing attack risk through adaptive privacy-personalization balance. Simultaneously, it facilitates communication overhead, which is the essential trade-off between privacy protection, accuracy, and resource efficiency for effective edge deployment.