Table 3 Computational complexity analysis of each component in the Architecture.

From: Privacy-preserving cyberthreat detection in decentralized social media with federated cross-modal graph transformers

Component

Time Complexity

Space Complexity

Deployment Notes

Text Feature Extractor

O(Ltxt·dtxt)

O(Vtxt·dtxt)

Efficient, on-device transformer

Image Feature Extractor

O(Pimg·Cimg²)

O(Pimg·dimg)

Lightweight CNN backbone

Audio Feature Extractor

O(Taud·Faud)

O(Faud·daud)

Uses temporal CNN, streaming capable

Graph Aggregation Layer

O(N·d2)

O(N·d)

Scalable with sparse adjacency

Cross-Modal Fusion

O(d2·M)

O(d·M)

Attention-based, M modalities

Transformer Layer

O(L2·d)

O(L·d)

Scales with sequence length, parallelizable

Federated Training Round

O(E·B·Clocal)

O(d)

E: epochs, B: batch, C: comm. steps

Privacy Mechanism (DP)

O(d)

O(d)

Minimal overhead, tunable