4.3.1 Three Core Technologies

The "Privacy Protection Triangle" with Different Strengths

1. Homomorphic Encryption: "Direct Arithmetic" on Encrypted DataThe core value of homomorphic encryption lies in "computing on data without decryption", which fundamentally avoids the risk of "plaintext exposure" of data during the computing process. Anonymous Imperium adopts a flexible adaptation strategy of "Fully Homomorphic Encryption (FHE) + Partially Homomorphic Encryption (PHE)":

• Technical characteristics:

  • Fully Homomorphic Encryption (optimized based on BGV algorithm): Supports all complex calculations such as addition, subtraction, multiplication, division, and logical operations. The calculation results of encrypted data are completely consistent with those of plaintext data, suitable for scenarios requiring in-depth data analysis (such as financial risk control model training).

  • Partially Homomorphic Encryption (focusing on additive homomorphism): Only supports addition and scalar multiplication operations. Although its functions are limited, its performance is 500% higher than that of fully homomorphic encryption, and it can run smoothly on ordinary terminal devices (mobile phones, laptops).

• Application scenarios:

  • Financial data analysis: Banks can directly perform "credit risk scoring" calculations on encrypted data without decrypting user transaction data, and output users' anonymous credit ratings, avoiding the risk of "data leakage caused by centralized storage".

  • Medical data statistics: Multiple hospitals jointly count the incidence rate of a certain disease. Through partially homomorphic encryption, the encrypted data of each hospital is aggregated and directly summed to obtain the regional total incidence rate, without sharing patients' personal medical records.

Core advantage: No third party is required during the computing process, and data ownership and control always belong to the user, suitable for high-sensitivity scenarios where "data is not easy to circulate".

1. Secure Multi-Party Computation (MPC): A "Privacy Calculator" for Decentralized CollaborationWhen computing tasks require multi-party participation (such as cross-enterprise data collaboration), Secure Multi-Party Computation (MPC) enables all parties to complete collaborative computing without "disclosing their own data" through "data sharding + decentralized computing".

Anonymous Imperium adopts a "threshold MPC + decentralized key management" solution:

• Technical characteristics:

  • Data sharding encryption: The computing task is decomposed into multiple subtasks. The input data of each party is cut into encrypted shards and distributed to participating nodes (such as DePIN computing power nodes) through encrypted channels.

  • Decentralized collaborative computing: Each node only processes the shard data it receives, exchanges intermediate computing results through an encryption protocol, and finally aggregates to obtain a unified answer. A single node cannot infer the original data.

  • Threshold signature and key management: A "2/3 threshold signature" mechanism is adopted. The computing result is valid only after being signed and confirmed by at least 2/3 of the participating nodes, avoiding the "computing failure caused by a single node's betrayal or attack".

• Application scenarios:

  • Cross-border e-commerce price comparison: Multiple e-commerce platforms jointly calculate the "regional average price" of a certain type of product. Each platform only uploads the encrypted shards of its own product prices, and obtains the average price through MPC collaboration, but cannot know the specific pricing strategies of other platforms.

  • Joint risk control modeling: Banks, payment institutions, and e-commerce platforms jointly build an "anti-fraud model". Each party provides its own user behavior data shards, and completes model training through MPC, which not only integrates multi-dimensional data to improve model accuracy, but also protects the privacy of their respective users.

Core advantage: Supports multi-party data collaboration, and all participants are in an equal position without a "data center node", completely eliminating the risks of single-point failure and data monopoly.

1. Federated Learning: "Data Does Not Move, Model Moves" in AI TrainingTo address the pain point of "data being scattered everywhere and difficult to centralize" in AI model training, the Aegis AI engine of Anonymous Imperium adopts an "improved federated learning architecture", enabling the model to complete collaborative optimization without "data migration":

• Technical characteristics:

  • Decentralized training: Original data is always stored on the local devices of data owners (such as enterprise servers and user terminals). Each participant trains the model on their own data locally, and only uploads "model parameter updates" (not the original data) to Aegis AI Core.

  • Secure parameter aggregation: Aegis AI Core integrates the model parameters of all participants through an "encrypted aggregation protocol" (based on secret sharing technology) to generate a globally optimized model, which is then distributed to each participant.

  • Differential privacy enhancement: "Carefully calibrated noise" is added to the model parameter updates. Even if an attacker infers through the parameters, they cannot accurately locate the original data of a certain user, further strengthening privacy protection.

• Application scenarios:

  • Medical AI diagnosis: Multiple hospitals across the country jointly train a "lung cancer image diagnosis model". Each hospital trains the model on its own case data locally and only uploads parameter updates. The final global model is far more accurate than the model trained with data from a single hospital, and patients' medical records never leave the hospital.

  • Anonymous credit scoring: User transaction data and node operation data in the ecosystem are scattered on their respective devices. Through federated learning, the "anonymous credit score model" is trained locally. After parameter aggregation, a unified scoring standard is generated. Users can obtain a credit score without exposing personal data.

Core advantage: Perfectly adapts to the needs of "local data storage", is compatible with large-scale AI collaborative training across devices and organizations, and takes into account model performance and data privacy.