CEICS brings AI inference to school-zone gateways — enabling real-time anomaly detection, privacy-preserving federated learning, and autonomous operation in conflict zones where cloud connectivity fails.
Cloud-centric safety systems fail when children need them most.
In conflict-affected regions like the Lake Chad Basin, over 800 school attacks occurred in a single year. The systems meant to protect children rely on infrastructure that gets cut first.
01
Latency too high for emergencies
Round-trip cloud latency of 250–450 ms exceeds the response window for pediatric seizures. Every 100 ms of delay is time lost for first responders.
02
Biometric data exposed in transit
Continuous uplink of raw heartbeat, GPS, and sensor data creates a persistent, large attack surface — precisely the data adversaries want about vulnerable children.
03
Brittle under network disruption
Centralized architectures go blind when connectivity drops — which is exactly when attacks happen. A system that fails during a crisis is worse than no system at all.
04
Model scarcity for rare events
A single school sees ~2 seizure events per month — not enough to train a reliable detector. Isolated edge nodes lack the data diversity needed for generalizable AI.
✗ Old cloud approach
All raw data leaves the school — serious privacy risk
268 ms average delay — exceeds seizure response threshold
Internet outage = system blind during attack
Single point of failure and data concentration
vs
✓ CEICS approach
Raw data never leaves the wearable — privacy by design
127 ms total latency — well within clinical threshold
Autonomous offline operation during disruption
10-node federation shares learning, not data
// 02 — Architecture
Four tiers. Each independent. Each essential.
CEICS is organized so that the failure of any higher tier degrades — but never disables — the safety guarantees of lower tiers. The school gateway keeps alerting even when the regional server is offline.
T1
Perception layer — wearable device
ESP32-C3 on each child. Measures GPS (2.5m CEP), 6-axis IMU at 50Hz, HR/SpO₂ via PPG, and EDA. Runs quantized INT8 CNN-LSTM inference every 200ms. Raw sensor values are discarded after inference and never transmitted.
22ms inference
T2
Edge intelligence layer — school gateway
NVIDIA Jetson Nano aggregates 200 wearables. Three concurrent threads: inference manager (full-precision CNN-LSTM with 2-window hysteresis), FL training manager (off-peak, 22:00–05:00), and alert/compliance manager.
127ms total
T3
Communication layer — adaptive dual-protocol
Finite-state machine with 4 states: LoRaWAN primary → 5G alert → hybrid degraded → satellite fallback. State transitions driven by RSSI and packet error rate every 30s. Emergency alerts always routed via highest available QoS.
auto-failover
T4
Analytics, privacy & compliance layer — regional server
Coordinates FL with Paillier homomorphic encryption. Global model updated in ciphertext space. 2-of-3 distributed key for decryption. PoA blockchain records FL rounds, alerts, and admin actions with 100% audit completeness.
2048-bit HE
// 03 — Interactive Simulation
Explore the system performance.
Based on NS-3 v3.38 + TensorFlow Federated 0.55 simulation across 150 nodes, 10 gateways. All metrics are mean ± SD across 5 independent runs (seeds 42–46).
SELECT A MODULE →
FINAL ACCURACY
95.3%
±0.41% · 95% CI
CONVERGENCE
200
rounds · Δacc <0.3%
GAP vs CENTRALISED
0.4 pp
p=0.09 non-significant
COMM COST SAVED
73%
486 MB vs 1,800 MB raw
Accuracy over 200 FL rounds
10
0.5
CEICSStandard FLEdge-onlyCentralised
Per-class precision / recall / F1
ACTIVE WEARABLES
150
10 gateways
ALERTS FIRED
0
Press start
AVG CONFIDENCE
—
P(anomaly)
FALSE ALERT RATE
2.1%
vs 4.8% cloud
▶ Live Alert Simulation
Streams real-time anomaly detection events from 150 wearable nodes across 10 school gateways.
Live alert stream
Press "Start simulation" above to begin →
Class distribution (live)
C0 NormalC1 FallC2 SeizureC3 ExitC4 Distress
C2 (Seizure) threshold sensitivity — drag to explore recall/precision tradeoff
0.92
Latency & Bandwidth comparison — four architectures measured across 5 simulation runs. Scroll down for the latency decomposition and ablation study table.
A step-by-step visual walkthrough of CEICS — no prior knowledge needed. Click through all 7 steps.
Step 1 of 7
Imagine a school in northeast Nigeria — in a conflict zone.
Schools in places like the Lake Chad Basin face real danger. Over 800 schools were attacked in one year. Children need protection — but what happens when the internet goes down during an attack?
!
Internet blackouts during attacks
Hostile groups deliberately cut communications infrastructure. Cloud-dependent systems go dark exactly when needed most.
!
Children carry sensitive biometric data
GPS, heartbeat, movement — if this data is intercepted, it creates a detailed map of where children are and how they're feeling.
!
Medical emergencies don't wait for cloud round-trips
A child having a seizure needs detection and response within minutes. 268 ms cloud latency may not sound long — but it's just the detection step.
The question CEICS answers: Can we build a safety system that works offline, protects privacy, detects emergencies in under 150ms, and gets smarter over time — all without ever exposing a child's data?
The old way sends everything to the cloud. That's the problem.
Before CEICS, most IoT child safety systems followed the same pattern: collect sensor data → send it all to a remote server → get back a decision. This introduces three fatal flaws.
🐢
Too slow
268 ms average. For a seizure, the 2–5 minute response window starts counting from the first twitch — not from when the cloud finally replies.
🔓
Too exposed
Raw biometric streams travel over public networks. Each packet contains identifiable data about a specific child's body and location.
💀
Too fragile
When the network goes down — which it does during attacks — the cloud system simply stops working. Zero offline capability.
The Ogundele et al. (2021) IoT child tracker achieved tracking at low cost but at 380–450 ms latency with unresolved data leakage. CEICS reduces latency by 66% and eliminates the leakage attack surface entirely.
The big idea: bring the brain to the school, not the data to the cloud.
Instead of uploading raw sensor data for processing far away, CEICS installs a small AI at the school itself. The AI thinks locally and immediately. Only the result — "seizure detected" — travels anywhere. And even then, it's encrypted.
👶 Child wearable
heartbeat · GPS · motion
👶 Child wearable
heartbeat · GPS · motion
+ 148 more
→signal only→←🔒 encrypted←
Edge gateway at school
AI thinks here · 22ms · Alerts in 127ms total · No raw data leaves
Why this works: The AI model on the gateway is trained across all 10 schools simultaneously — without any school sharing its children's data. Schools share what they learned, not what they saw. This is federated learning.
CEICS has 4 layers — like a team where each person has one job.
Each tier has a specific responsibility. If the top tier fails, the lower ones keep working independently.
1
Wearable — on the child
Smart badge/bracelet on ESP32-C3. Measures heartbeat, movement, location, skin response. Runs a mini-AI that makes a decision every 200ms. Raw values are never stored or transmitted.
2
Edge gateway — at the school
Small computer (Jetson Nano) that receives signals from all 150 wearables. Runs the full AI decision. Issues alerts in 127ms. Trains the model at night. Works offline.
3
Communication — the radio system
Uses LoRaWAN (long-range low-power) and 5G. Automatically switches between them based on signal quality. Falls back to satellite if both fail. Emergencies always use the best available channel.
4
Regional server — the coordinator
Never sees raw data. Receives only encrypted math updates from each school gateway. Combines them to make every school's AI smarter. Records every event permanently on a tamper-proof blockchain.
The AI watches for 5 types of dangerous events — every 5 seconds.
Using heartbeat, movement, location, and skin data simultaneously, the model classifies each 5-second window into one of five categories.
●
C0 — Normal
Child is fine. Walking, sitting, running normally. Detected with 97% accuracy.
●
C1 — Fall or assault
Sudden violent impact detected by accelerometer. Triggers alert to nearest teacher or guard.
GPS outside school boundary for >30 seconds. Could mean abduction or wandering.
●
C4 — Body distress
Abnormal heartbeat, low oxygen, or high stress. Could be shock, fear, or medical emergency.
⏱
127 ms total
Two consecutive windows must agree before alert fires. Prevents false alarms from sudden movements.
The two-window hysteresis filter means a child who jumps up and knocks their badge won't trigger a fall alert. But a genuine fall that persists into the second window definitely will.
How 10 schools teach each other — without sharing a single child's data.
This is the most sophisticated part. Each school's AI learns from its own children. But instead of sending raw data to improve, it sends only the abstract mathematical "lesson" — encrypted so heavily that even the receiving server can't read it.
1
School trains locally at night
Between 22:00–05:00, the gateway trains on its own children's data from the past 72 hours. No outside connection needed for this step.
2
Sends only the "lesson" — encrypted
The gateway computes what it learned (the gradient update) and encrypts it with 2048-bit Paillier encryption before sending. It looks like random numbers to anyone intercepting it.
3
Server combines lessons from all schools — still encrypted
The regional server adds up all 10 schools' encrypted updates mathematically. Paillier encryption has a special property: you can add encrypted numbers without decrypting them first.
4
Improved model sent back to all schools
The combined lesson is decrypted using a 2-of-3 key split — no single person can decrypt alone. Every school gets a smarter model. The cycle repeats every 24 hours.
After 200 rounds, CEICS achieves 95.3% accuracy — within 0.4% of what you'd get if all the raw data had been pooled together (95.7%) — while maintaining complete data locality.
What this actually means for children in real danger.
These numbers aren't just research metrics. Each one represents a real outcome for a real child in a real school.
127 ms
Alert latency
vs 268 ms cloud. In the 2–5 min seizure window, this difference could mean a child gets help in time.
99.1%
Seizure recall
Misses fewer than 1 in 100 seizures. Better than the best specialized single-class detector.
−65%
Data over the air
Less to intercept. Works in low-bandwidth zones. Less power used by wearables.
78%
Attack resistance
Hackers attempting to steal data have 78% of attempts thwarted vs 22% for cloud systems.
100%
Audit trail
Every alert recorded on tamper-proof blockchain. Schools can prove what happened to authorities.
Offline
Conflict resilience
Gateway keeps detecting and alerting locally when internet is cut during an attack.
The bigger picture: CEICS is a blueprint for the Oslo Safe Schools Declaration — proving that the gap between international policy commitments and operational reality can be bridged through principled, privacy-respecting engineering. Any school in any conflict zone can use this.
// 05 — Key Results
Statistically significant across every dimension.
Five independent simulation runs (seeds 42–46). One-way ANOVA F(3,16) = 847.3, p < 0.001, η² = 0.994. All baseline comparisons significant at p < 0.0001.
95.3%
Anomaly detection accuracy
±0.41% across 5 runs. 95% CI [94.89%, 95.71%]. Model type explains 99.4% of variance.
+6.9 pp vs cloud · d=16.8
127 ms
End-to-end alert latency
±3.2 ms. 52.6% improvement over 268ms cloud baseline. Well within 150ms clinical threshold.
d=21.6 · p<0.0001
99.1%
Seizure class (C2) recall
±0.41%. Statistically equivalent to the best specialized single-class seizure detector (IIETA 2025: 99.0%, p=0.65) while simultaneously performing 4 other detection tasks.
+7.7 pp vs Masci 2025
−65%
Network bandwidth reduction
CEICS consumes 35% of cloud baseline bandwidth through on-device extraction, LoRaWAN duty-cycling, and FL gradient sharing.
+22% energy efficiency
78.4%
Intrusion resistance rate
+56.3 pp over cloud baseline (22.1%). 97.2% resistance to passive eavesdropping. 63.7% FL replay resistance vs 43.4% standard FL.
d=28.2 · p<0.0001
0.4 pp
Federated convergence gap
Gap between federated (95.3%) and centralised (95.7%) training at round 200. Non-significant: Welch t(7.2)=1.94, p=0.09. Full data locality maintained.
200 rounds · 486 MB total
// 06 — The Paper
Collaborative Edge Intelligence for Secure Child Location Monitoring in IoT Systems
A Privacy-Preserving Framework Aligned with the Oslo Safe Schools Declaration
Abstract
The proliferation of Internet of Things (IoT) devices in educational environments has generated significant demand for real-time child location monitoring; however, prevailing cloud-centric architectures impose prohibitive latency, consume excessive network bandwidth, and expose sensitive biometric data to interception risk. This paper presents the Collaborative Edge-Intelligent Computing System for Child Safety (CEICS), a four-tier privacy-preserving framework that relocates AI inference to school-zone edge nodes. On a combined dataset of 10,847 multi-modal sensor windows drawn from the MIT-BIH Arrhythmia Database and SHAR-100-20, CEICS achieved mean anomaly detection accuracy of 95.3% ± 0.41% (95% CI: [94.89%, 95.71%]) and macro F1-score of 94.7% ± 0.38%, statistically significantly outperforming all three baselines (p < 0.001). End-to-end emergency alert latency was reduced to 127 ± 3.2 ms — a 52.6% improvement over the cloud baseline.
Authors
Sylvester Oga Ogaji
School of Computer and Information Sciences, Indira Gandhi National Open University, India
Oluwagbemisola J. Akinlade-Ogaji
Department of Public Health, Ladoke Akintola University of Technology, Nigeria
Adeniji Samuel Olamilekan
Department of Computer Science, Crown Polytechnic, Ado Ekiti, Nigeria
Adeyanju Emmanuel Abayomi
Department of Computer Science Education, National Open University of Nigeria
Alabi Ayomide Praise
Department of Mass Communication, Kwara State University, Malete, Nigeria
Sylvester Oga Ogaji is the principal investigator and lead architect of the CEICS framework — the first IoT system to jointly optimise latency, detection accuracy, privacy, and communication resilience for child safety monitoring in Nigeria's conflict-affected school zones. His work sits at the intersection of edge computing, privacy-preserving machine learning, and humanitarian technology, motivated by the documented crisis of school attacks across the North-West, North-Central, and North-East corridors.
CEICS
Flagship research framework
Collaborative Edge-Intelligent Computing System for Child Safety — IoT wearable + federated CNN-LSTM + PoA blockchain
IGNOU
Doctoral research institution
School of Computer & Information Sciences · Computer Science · DSR methodology
JV ImpactVR
Nigerian research initiative
JV ImpactVR Initiative — research, technology transfer, and humanitarian innovation in Nigeria
Oslo SSD
Policy alignment
CEICS is the first IoT framework designed with explicit Safe Schools Declaration compliance built into the architecture