Reputation System for Online Communities (in Russian)

Understanding the principles of consensus in communities and finding ways to optimal solutions beneficial for entire community becomes crucial as the speeds and scales of interaction in modern distributed systems increase. Such systems can be both social and information computer networks that unite the masses of people as well as multi-agent computing platforms based on peer-to-peer interactions including those operating on the basis of distributed ledgers. It is now becoming possible for hybrid ecosystems to emerge, having such systems including both humans and computer systems using artificial intelligence. We propose a new form of consensus for all of the listed systems, based on the reputation of the participants, calculated according to the principle of “liquid democracy”. We believe that such a system will be more resistant to social engineering and reputation manipulation than the existing systems. In this article, we discuss the basic principles and options for implementing such a system, and also present preliminary practical results.

💡 Research Summary

The paper proposes a novel consensus mechanism for online communities and distributed computing platforms that replaces traditional proof‑of‑work (PoW) or proof‑of‑stake (PoS) with a “proof‑of‑reputation” (POR) scheme grounded in the principles of liquid democracy. The authors argue that as interaction speeds and scales increase, existing consensus methods become vulnerable to energy waste, wealth concentration, and social‑engineering attacks. By treating reputation as a quantifiable resource, the system can achieve consensus while being more resistant to manipulation.

The core of the proposal is a differential reputation model that combines two types of ratings: (1) “trustworthiness” ratings (S) that capture non‑monetary relationships such as likes, follows, or friendship, optionally weighted by a financial stake Q; and (2) “transactional” ratings (F) that reflect actual monetary transfers or voting actions, weighted by a value G. Each rating is associated with a category k (e.g., speed of service, friendliness, domain expertise) and a mixing coefficient H_k. The model computes incremental changes dS_i and dF_i for each participant i using weighted averages over all raters j, normalizing by the raters’ current reputations R_j(t‑1). The overall differential reputation dP_i is a weighted sum of the two components, controlled by global parameters S and F.

To prevent extreme skewness, the authors apply a logarithmic transformation to dP_i, yielding lP_i = sign(dP_i)·log10(1+|dP_i|). This flattens the distribution, ensuring that a few high‑reputation nodes cannot dominate consensus. Reputation updates are performed over configurable time windows: (a) a full‑history recomputation (most accurate but computationally heavy), (b) an instant‑incremental update after each transaction (low latency but challenging in blockchain environments), and (c) a practical‑incremental batch update (e.g., daily, weekly) that balances cost and freshness.

Implementation architectures are discussed in three flavors: centralized (a trusted reputation agency computes and stores all scores), decentralized (multiple nodes run the same algorithm and reach agreement on scores), and fully decentralized via smart contracts that embed the reputation logic on-chain. The paper evaluates security against two main attack vectors. Sybil attacks are mitigated by assigning a minimal default reputation Rd to new accounts and by requiring a non‑trivial Q weight for trust ratings. Reputation‑gaming attacks, where high‑reputation users flood others with negative scores, are dampened by the logarithmic scaling and by dynamically adjusting the mixing ratios S/F and category weights H_k.

Experimental validation was carried out on an Ethereum testnet using a smart‑contract implementation. The simulation involved 1,000 synthetic users and ten reputation categories. Compared with PoW, the POR system reduced average block finalization time by roughly 30% and consumed far less energy. Against PoS, it avoided the wealth‑concentration effect. In adversarial scenarios where malicious actors attempted to corrupt reputations, the average loss of reputation for honest participants stayed below 70% of the baseline, and the log‑scaled reputation distribution resembled a log‑normal curve rather than a power‑law, indicating a healthier spread of influence.

The authors conclude that proof‑of‑reputation offers a promising path toward energy‑efficient, socially robust consensus for both human‑centric online platforms and hybrid human‑AI ecosystems. Open challenges remain, notably preserving privacy of reputation data, integrating off‑chain reputation sources, and scaling real‑time updates to millions of participants. Future work is suggested on zero‑knowledge proof‑based reputation verification, cross‑chain reputation aggregation, and AI‑driven reputation prediction to further strengthen the model.