Proxy Aggregator: Indirect Data Aggregation

• Proxy aggregators are systems that collect and reconcile proxy representations, enabling efficient routing, validation, and decision-making across varied technical domains.
• They are applied in areas such as Ethereum smart contracts, online experiments, and federated learning to score reliability, classify behaviors, and preserve privacy.
• By exposing hidden relationships and vulnerabilities like function collisions or fragility, proxy aggregators offer actionable insights for enhancing system performance and security.

Searching arXiv for the cited papers to ground the article and confirm metadata. Search query: "Proxy Aggregator arXiv proxy contracts PROXIMA PlexiChain open proxies proxy server federated proxy model" A proxy aggregator is a system that collects, reconciles, and operationalizes proxies rather than acting directly on the underlying object. In contemporary research, the term does not denote a single standardized artifact. It appears in at least six technical senses: chain-scale analyzers for proxy smart contracts, reliability scorers for proxy metrics, federated-learning schemes that aggregate through proxy models, infrastructures that broker or schedule network/API proxies, market coordinators that virtualize the aggregator role through smart contracts or agentic interfaces, and algorithms that resolve proxy-based delegation graphs (Chen et al., 2024, Amudala, 15 Apr 2026, Kalra et al., 2021).

1. Semantic scope and unifying idea

Across these literatures, the common structure is indirect aggregation. The aggregated object is not the final outcome itself, but a proxy representation that is easier to observe, route, validate, or exchange. This suggests a family resemblance rather than a single formal definition.

Domain	Proxy object	Aggregator function
Ethereum	Proxy contracts and logic links	Discover, classify, analyze collisions
Online experiments	Proxy metrics	Score reliability for launch decisions
Federated learning	Proxy models or proxy weights	Exchange and fuse knowledge
Networking and APIs	Open proxies or token-backed workers	Route, throttle, and monitor requests
Energy and delegation	Flexibility offers or delegated votes	Clear, settle, or resolve weighted flow

The strongest technical instantiations share three properties. First, they represent a large population through a smaller or more tractable proxy layer. Second, they aggregate across relationships rather than isolated objects: proxy–logic pairs, experiment–segment pairs, client–proxy links, or delegation edges. Third, they usually expose system-level pathologies that are not visible locally, such as selector collisions, segment fragility, update interference, insider leakage, or cyclic delegation (Ebrahimi et al., 1 Jan 2025, Nasr et al., 2017, Degrave, 2014).

2. Proxy aggregation in Ethereum smart contracts

In Ethereum, the proxy design pattern splits an original contract into a proxy contract containing storage and a logic contract containing implementation logic. Calls are forwarded through DELEGATECALL, so the logic executes in the proxy’s storage context. This architecture enables upgradeability, but it also creates function-collision and storage-collision attack surfaces (Chen et al., 2024).

The most explicit ecosystem-scale proxy aggregator in this literature is Proxion. It is described as “an automated cross-contract analyzer that aims to uncover all proxy smart contracts in Ethereum,” and its distinguishing feature is coverage of hidden contracts with neither verified source code nor past transactions. Applied to over 36 million alive contracts from 2015 to 2023, it identified 19,599,317 proxy contracts, or 54.2% of alive contracts, and about 1.5 million contracts with at least one collision issue. Its pipeline combines bytecode filtering for DELEGATECALL, EVM emulation with crafted calldata, recovery of logic addresses from the stack, storage-slot reconstruction through eth_getStorageAt, and cross-contract collision analysis (Chen et al., 2024).

A second line of work treats proxy aggregation as systematized measurement of proxy prevalence and type. A large-scale exploratory study on 50,845,833 contracts and 1,695,517,186 transactions reports that 14.2% of all deployed contracts are active proxy contracts, that proxies are more actively used than non-proxies, that 67.8% of sampled proxies act as forwarders or interceptors, and that 32.2% enable upgradeability. In the same study, 79% of proxies matched known reference implementations and 29.4% were ERC-1167 minimal proxies (Ebrahimi et al., 1 Jan 2025). The prevalence estimates differ from Proxion because the datasets and detection criteria differ: one study targets active proxies through traces, whereas the other analyzes alive contracts with hidden-proxy recovery.

The pattern space itself has been cataloged in a survey of upgradeable smart contracts. That survey organizes inherited storage, eternal storage, unstructured storage, UUPS/EIP-1822, Transparent proxies, EIP-1538, EIP-2535 Diamonds, Beacon proxies, Registry proxies, EIP-1167 minimal proxies, and EIP-897 DelegateProxy into a single taxonomy. Using a Slither-based reimplementation, it reports that approximately 70 percent of contracts initially flagged as upgradeable proxies were false positives and were eliminated by more robust detection (Meisami et al., 2023). In this sense, a smart-contract proxy aggregator is simultaneously a registry, a classifier, and a security scanner.

3. Proxy aggregation for decision-making under imperfect observability

In online experimentation, proxy metrics are short-horizon signals used in place of slow-moving Overall Evaluation Criteria. PROXIMA formalizes a proxy aggregator for this setting by defining a composite reliability score

$$R = w_C \cdot C + w_{\text{DA}} \cdot \text{DA} + w_{\text{FR}} \cdot (1-\text{FR}),$$

where $C$ is normalized effect correlation, DA is directional accuracy, and FR is segment-level fragility rate. The framework does not predict long-term treatment effects numerically; it audits whether a candidate proxy leads to correct ship/no-ship decisions and where it fails (Amudala, 15 Apr 2026).

The empirical results make the aggregation role concrete. On Criteo, early engagement metrics achieve a composite reliability of 0.80 and 98.4% average decision agreement with an oracle policy; on KuaiRec, they achieve 0.62. Fragility analysis shows 13% fragility in advertising and 68% fragility in recommendation, while directional accuracy remains above 96% in both domains (Amudala, 15 Apr 2026). The key point is that a proxy aggregator here summarizes three distinct failure modes at once: magnitude alignment, sign agreement, and segment robustness.

A related but economically different use of proxy aggregation appears in decision markets. “Proxy Forecasting to Avoid Stochastic Decision Rules in Decision Markets” replaces outcome-based scoring with proxy-based scoring so that the principal can use a deterministic decision rule. The paper proposes three mechanisms: one where the principal uses her own signal as the proxy, one where a partial-profit contract aligns one agent and retrieves its signal as proxy, and one where proper peer prediction elicits signals from two agents. The mechanism then aggregates forecasts about the proxy and maps the aggregated forecasts and the proxy to a deterministic action; in a multi-agent bandit system, the first mechanism reaches performance similar to a Bayesian inference model with access to all information held by the agents (Wang et al., 2023).

These two literatures use “proxy aggregator” differently, but both make the same technical move: they separate the decision problem from the directly observed long-run outcome, and instead aggregate a structured proxy object with explicit diagnostics or incentive guarantees.

4. Proxy aggregation in federated and decentralized learning

In machine learning, proxy aggregation is used to exchange information without directly sharing the primary model or raw data. “Decentralized Federated Learning through Proxy Model Sharing” introduces ProxyFL, in which each participant maintains a private model and a publicly shared proxy model. The proxy models are averaged across a directed graph with PushSum, while private models remain local. This removes the centralized server, allows model heterogeneity, and places differential privacy noise on the shared proxy rather than on the private model (Kalra et al., 2021).

A second ProxyFL, for federated semi-supervised learning, treats the learnable weights of the classifier as class proxies. On the server, Global Proxy Tuning optimizes global proxies against outliers instead of directly averaging full weights; on the client, Indecisive-Categories Proxy Learning re-includes low-confidence unlabeled samples through a positive-negative proxy pool. The framework is explicitly designed to mitigate both external heterogeneity across clients and internal heterogeneity between labeled and unlabeled data within a client (Chen et al., 24 Feb 2026).

FedProxy extends the same pattern to LLM adaptation. It compresses a proprietary LLM into a Proxy Small LLM, performs federated optimization on that proxy with heterogeneity-aware aggregation, and then fuses the learned proxy back into the full LLM by a training-free plug-in mechanism. In the heterogeneous LLaMA2-7B, 50% compression setting, the reported scores are QA 0.6585 and GLUE 0.8861 for centralized fine-tuning, QA 0.4750 and GLUE 0.5933 for FedOT, and QA 0.6011 and GLUE 0.7944 for FedProxy (Fan et al., 21 Apr 2026).

A plausible synthesis is that ML proxy aggregators replace impossible or undesirable direct aggregation with a shared surrogate space. In decentralized FL that surrogate is a public proxy model; in FSSL it is class-proxy weights; in federated LLM tuning it is a compressed proxy SLM. The common technical purpose is to preserve privacy or IP while still supporting cross-client aggregation.

5. Network, API, and censorship proxy infrastructures

In networking, a proxy aggregator often means a service that collects or schedules access to proxy endpoints. The open-proxy measurement literature studies this usage directly. Over 50 days, one study collected 107,034 unique open proxies from aggregator sites and issued about 13 million proxy requests. More than 92% of listed open proxies were unresponsive. Among HTTPS-supporting proxies, 1.06% performed TLS man-in-the-middle attacks; among proxies connecting to a logging server, 66.08% leaked the client IP in request headers; and HTML, executable, and other content were sometimes modified to inject cryptojacking code, remote access trojans, and other malware (Mani et al., 2018). In this literature, the proxy aggregator site is the entry point to a brittle and frequently hostile ecosystem.

A more controlled form of infrastructure aggregation appears in “GitHub Proxy Server.” There the proxy aggregator sits between client applications and the GitHub REST or GraphQL APIs, creates one worker per token, chooses the worker with the highest capacity and availability, and enforces request intervals, timeouts, and minimum remaining quota. In the evaluation, total execution time improved from 35 minutes 22 seconds to 32 minutes 13 seconds with one token, and from 36 minutes 55 seconds to 26 minutes 49 seconds with three tokens (Borges et al., 23 May 2025). The proxy is not aggregating GitHub data semantically; it is aggregating token budgets and request scheduling.

The censorship-circumvention literature uses the term differently again. “Enemy At the Gateways” models proxy distribution in systems such as Tor, Psiphon, and Lantern as a game between the proxy distributor and a censor deploying insider identities to discover and block proxies. The paper derives an optimal proxy assignment mechanism against an optimal censorship adversary and evaluates it through simulations (Nasr et al., 2017). Here the proxy aggregator is a distributor whose task is not to measure or schedule proxies, but to allocate secrecy under adversarial leakage.

Taken together, these works show that network-oriented proxy aggregators are defined as much by threat model as by routing logic. Some are scheduling layers over scarce credentials, some are measurement front ends over untrusted endpoints, and some are defensive allocation mechanisms against strategic compromise.

6. Intermediation, settlement, and delegated resolution

In energy systems, a proxy aggregator is often an intermediary that bundles and clears flexibility on behalf of many participants. PlexiChain recasts this role as a permissioned blockchain framework in which registration, bidding, matching, verification, and settlement are implemented by smart contracts. Flexible Resource Owners submit offers $O_i = (F_i, \pi_i, \text{constraints}_i)$, the clearing contract aggregates offers against a request $R$, and settlement uses delivered flexibility and penalties through formulas such as

$$\Delta F_i(t) = B_i(t) - M_i(t), \qquad r_i(t) = \pi^\ast \cdot \Delta F_i(t) - \lambda \cdot |\epsilon_i(t)|.$$

The paper reports robustness against MadIoT and False Data Injection attacks and lower computation and communication costs than other blockchain-based applications in resource-constrained environments (Karumba et al., 2022).

Conversational Demand Response relocates the same intermediary function into an agentic-AI architecture. An aggregator agent receives a system-level flexibility request, selects households, dispatches natural-language DR events, and updates its portfolio after each HEMS response. Household feasibility is evaluated locally through an optimization-based battery tool. In the proof-of-concept evaluation, all interactions complete in under 12 seconds (Makroum et al., 6 Mar 2026). The aggregator here is a proxy in the literal sense of representing prosumers upstream while delegating household-level optimization to local HEMS agents.

Liquid-democracy research provides a mathematically explicit aggregation problem for proxy delegation. In multi-proxy transitive vote delegation, each voter can split a vote across several delegates, and the proxy aggregator must resolve the resulting graph. The paper models the final vote-weight vector $S$ by

$S = J + A S, \qquad S = (I-A)^{-1} J,$

where $J$ is the vector of base votes and $A$ is the delegation matrix. This formulation resolves branching, recombination, and cycles in a single linear system (Degrave, 2014). Unlike the previous examples, the proxy aggregator is purely computational: it turns a delegation graph into effective voting weights.

These cases illustrate a broader point. In energy, conversational DR, and liquid democracy, the proxy aggregator is not merely a proxy list or a scorecard. It is an institution or algorithm that transforms delegated or representative signals into executable commitments, settlements, or vote weights.