Federated Learning

Figure 1: Federated learning architecture showing how InterLink Network trains AI models across distributed devices while preserving user privacy. Model updates are aggregated centrally while raw biometric data never leaves user devices.

Training Phase

InterLink Network employs federated learning (FL) to train its AI models (e.g., deepfake detection, fraud detection, etc.) across thousands of user devices and nodes without exposing raw biometric data. As a next-generation digital identity platform designed for the decentralized era, InterLink ID provides a secure and private method to verify one's unique human identity, acting as a universal passport in the digital world. Unlike conventional federated learning systems, InterLink first establishes user uniqueness through its proprietary hashing technology, assigning each enrolled user a distinct cryptographic identity within the Network. InterLink ID also converts biometric data provided by users into secure, irreversible embeddings for federated training, ensuring privacy and security while enabling collaborative model improvement.

By using biometrics (a face scan) instead of passwords or documents, InterLink ID dramatically improves both security and convenience. Users can prove "I am a real person and I am unique" without sharing sensitive personal details each time. In a federated round, each client (user's device or a node) computes updates to the global model using only these secure irreversible high-dimensional embeddings of the provided user data, and only model updates (e.g., gradients or weight deltas) are sent to the aggregator (server), never the raw biometric data. The server then performs a weighted averaging of the updates to improve the global model. Formally, if $F_k(w)$ is the local loss on client $k$'s data, federated averaging optimizes the global model $w$ by minimizing the global loss $F(w)=\sum_{k=1}^{K}\frac{n_k}{N}F_k(w)$, where $n_k$ is the number of samples on client $k$ and $N$ is the total number of samples across all devices. After each round, model parameters are updated as:

$$w_{t+1} \leftarrow w_t - \eta \sum_{k=1}^{K} \frac{n_k}{N} \nabla F_k(w_t)$$

which is a gradient descent step combining all clients' contributions. Through this process, the AI model improves collectively: for example, learning to better distinguish between real transactions and fraudulent ones as more clients enroll in the system, without compromising individual biometric data. Federated learning comes with theoretical convergence guarantees under certain conditions, and techniques like secure aggregation and differential privacy can be layered in to ensure no single participant's data can be reconstructed from model updates.

Resource sharing and optimization: InterLink Network optimizes federated learning for diverse, heterogeneous devices through strategies like FedAvg with adaptive learning rate $\eta$ and lightweight model architectures. Techniques such as model pruning and model quantization reduce computational overhead, allowing even resource-constrained mobile devices to participate with minimal latency or power consumption. Users can opt-in to contribute resources during idle periods (e.g., while charging or on Wi-Fi), earning token rewards proportional to their computational effort. On-chain checkpoints, recorded via blockchain, enhance transparency and fault tolerance. This distributed computing framework not only boosts model accuracy over time but also eliminates centralized data vulnerabilities, aligning with InterLink ID's decentralized ethos.

Backup Mechanism for Intermediate Nodes: To ensure resilience, InterLink ID implements a robust backup mechanism for intermediate nodes that store encrypted biometric embeddings. As depicted in Figure 1, these nodes operate as a distributed layer between clients and the aggregator, enhancing security and continuity. The mechanism includes:

1. Redundant Storage: Encrypted embeddings are replicated across multiple nodes, providing fault tolerance.

2. Real-time Monitoring: Continuous health and security checks detect anomalies or failures in real time.

3. Automatic Failover: Upon detecting an attack or node failure, the system activates backup storage, seamlessly restoring encrypted data with minimal disruption.

4. Data Integrity Checks: Cryptographic hash functions (e.g., SHA-256) verify the consistency and authenticity of stored embeddings, ensuring resilience against tampering.

This architecture, illustrated in Figure 1, ensures that the federated learning process remains operational and secure, even under adversarial conditions, preserving both data integrity and user privacy.

Inference Phase

The inference phase in a federated learning system is crucial for deploying the trained model to make predictions on new data while maintaining the privacy and security principles established during training. In the context of InterLink ID, the inference phase involves using the globally trained model to verify user identities without exposing sensitive biometric data.

Inference Process:

1. Local Data Processing: When a user attempts to authenticate, their biometric data (e.g., a face scan) is processed locally on their device. Let $B$ represent the raw biometric data. The biometric data is converted into secure, irreversible embeddings $E$ using the same feature extraction and hashing techniques employed during training: $E = f(B)$ where $f$ is the feature extraction function.

2. Local Model Application: The locally processed embeddings $E$ are then fed into the globally trained model $M$, which resides on the user's device. This model has been updated through federated learning rounds and contains the collective knowledge from all participating devices: $P = M(E)$ where $P$ is the prediction output.

3. Prediction Generation: The model generates a prediction $P$ based on the local embeddings. For identity verification, this prediction could be a probability score $p$ indicating the likelihood that the user is who they claim to be: $p = \sigma(P)$ where $\sigma$ is the sigmoid activation function.

4. Secure Communication: If necessary, the prediction $p$ or a summary of the inference results can be securely communicated to a central server or a decentralized network for further validation. However, the raw biometric data $B$ and embeddings $E$ remain on the user's device, ensuring privacy.

5. Decision Making: Based on the prediction $p$, the system makes a decision regarding the user's authentication request. This decision can be made locally or in conjunction with additional verification steps performed by the central server or decentralized network.

Advantages of Federated Inference:

• Privacy Preservation: By keeping the raw biometric data $B$ and embeddings $E$ on the user's device, the inference phase maintains the privacy and security principles of federated learning.

• Reduced Latency: Local inference reduces the need for constant communication with a central server, resulting in faster authentication times.

• Scalability: The decentralized nature of federated inference allows the system to scale efficiently, handling a large number of authentication requests without overloading a central server.

• Robustness: The use of a globally trained model $M$ ensures that the system benefits from the collective knowledge of all participating devices, improving the accuracy and robustness of predictions.

Example Workflow:

1. User Authentication Request: A user initiates an authentication request by providing a biometric input $B$ (e.g., a face scan).

2. Local Processing: The user's device processes the biometric input $B$, converting it into secure embeddings $E$: $E = f(B)$

3. Local Inference: The locally stored model $M$ generates a prediction $P$ based on the embeddings $E$: $P = M(E)$

4. Secure Validation: If needed, the prediction $p = \sigma(P)$ is securely communicated to a central server or decentralized network for additional validation.

5. Authentication Decision: The system makes a final decision on the authentication request, granting or denying access based on the prediction $p$.

Figure 2: Example workflow of the federated inference process in InterLink ID.

By integrating federated learning into both the training and inference phases, InterLink ID ensures a secure, private, and efficient identity verification process that aligns with the principles of decentralization and user control.

Integration with Proprietary Technology: InterLink ID’s federated learning leverages its proprietary hashing technology to create cryptographic identities and secure embeddings. Each user’s biometric data is salted and hashed into a unique identifier, enabling precise tracking of contributions without revealing personal details. During training, embeddings are generated via a hybrid CNN-hashing pipeline, ensuring compatibility with the global model while thwarting reverse-engineering attempts. This integration enhances both privacy and the system’s ability to scale across diverse populations.

Performance Evaluation: InterLink ID’s federated learning system achieves a False Acceptance Rate (FAR) below 0.001 and a False Rejection Rate (FRR) below 0.005, validated across 10,000 devices. Compared to centralized models, it offers a 20% improvement in fraud detection accuracy due to its diverse training data. Inference latency averages 450 ms on mid-range smartphones, with energy consumption reduced by 30% through optimization techniques, ensuring accessibility and efficiency.

User Incentives and Participation: To encourage participation, InterLink ID offers token rewards based on computational contributions, calculated as $R_k=\alpha \cdot n_k \cdot t_k$, where $\alpha$ is a reward rate, $n_k$​ is the client’s sample size, and $t_k$ is training time. Users opt-in via a transparent interface, controlling when their device trains (e.g., overnight). This incentivized model has boosted participation rates by 40%, creating a self-sustaining network that enhances model accuracy over time.

Future Directions: InterLink ID aims to enhance federated learning with homomorphic encryption, enabling computations on encrypted updates for added privacy. Plans also include integrating secure multi-party computation (SMPC) to further decentralize aggregation, reducing reliance on a central server. These advancements will position InterLink ID as a leader in privacy-preserving AI, adapting to emerging threats like quantum attacks.