AdFL: In-Browser Federated Learning for Online Advertisement

Since most countries are coming up with online privacy regulations, such as GDPR in the EU, online publishers need to find a balance between revenue from targeted advertisement and user privacy. One way to be able to still show targeted ads, based on user personal and behavioral information, is to employ Federated Learning (FL), which performs distributed learning across users without sharing user raw data with other stakeholders in the publishing ecosystem. This paper presents AdFL, an FL framework that works in the browsers to learn user ad preferences. These preferences are aggregated in a global FL model, which is then used in the browsers to show more relevant ads to users. AdFL can work with any model that uses features available in the browser such as ad viewability, ad click-through, user dwell time on pages, and page content. The AdFL server runs at the publisher and coordinates the learning process for the users who browse pages on the publisher’s website. The AdFL prototype does not require the client to install any software, as it is built utilizing standard APIs available on most modern browsers. We built a proof-of-concept model for ad viewability prediction that runs on top of AdFL. We tested AdFL and the model with two non-overlapping datasets from a website with 40K visitors per day. The experiments demonstrate AdFL’s feasibility to capture the training information in the browser in a few milliseconds, show that the ad viewability prediction achieves up to 92.59% AUC, and indicate that utilizing differential privacy (DP) to safeguard local model parameters yields adequate performance, with only modest declines in comparison to the non-DP variant.

💡 Research Summary

AdFL introduces a novel in‑browser federated learning (FL) framework that enables online publishers to improve ad targeting while complying with stringent privacy regulations such as GDPR and CCPA. Unlike prior FL work that focuses on mobile devices or server‑client architectures, AdFL operates entirely within the web browser, requiring no additional software installation. The system consists of a publisher‑hosted FL server and a lightweight JavaScript client that runs in every visitor’s browser.

Key components of the client are: (1) a data‑collection module that leverages standard browser APIs (e.g., MutationObserver) to capture ad‑related events such as viewability timestamps, clicks, dwell time, and page‑content features; (2) a preprocessing pipeline that normalizes and formats these logs into feature vectors; (3) a TensorFlow.js‑based learning engine that trains a compact neural network (typically two to three dense layers) using only the data stored locally; (4) a synchronization layer that tracks the current global model version via first‑party cookies and asynchronously uploads model updates; and (5) a dynamic concatenation layer that activates only the features available on a given device, allowing the same model to run on heterogeneous browsers.

The server aggregates client updates using standard FedAvg or similar algorithms, producing a new global model that is redistributed to all clients. To protect privacy beyond the inherent data locality of FL, AdFL adds differential privacy (DP) noise to the transmitted model parameters, achieving an ε‑DP budget of roughly 1–2. Experiments on two non‑overlapping real‑world datasets (10‑day and 30‑day logs from a site with 40 K daily visitors) demonstrate that data collection and preprocessing incur less than 5 ms latency, while training and inference take under 3 ms on desktop browsers and under 7 ms on mobile browsers. The viewability prediction model reaches an AUC of 92.59 % without DP and only a modest drop to ~90.8 % with DP, confirming that privacy‑preserving noise does not significantly harm ad‑revenue potential.

Compared with related work, AdFL uniquely supports horizontal FL on browser‑available features, avoids the need for vertical FL that requires cross‑entity data sharing, and provides a complete end‑to‑end pipeline (collection, preprocessing, training, aggregation, inference) tailored to the online advertising ecosystem. Existing in‑browser FL frameworks either lack ad‑specific data handling, rely on synthetic datasets, or require peer‑to‑peer communication that is impractical for publishers.

The paper also discusses limitations: client heterogeneity (different browsers, hardware capabilities) can affect convergence speed; communication overhead from model updates remains non‑trivial; and the current system does not integrate directly with real‑time bidding engines, leaving the pathway from model prediction to ad auction decisions open. Future research directions include compressing updates to reduce bandwidth, extending the model to multitask learning (viewability, click‑through, conversion), enabling multi‑tenant FL across several publishers, designing APIs for seamless integration with demand‑side platforms, and developing adaptive mechanisms that maintain performance as user behavior evolves over long periods.

In summary, AdFL demonstrates the feasibility of deploying federated learning entirely within web browsers for privacy‑preserving ad targeting. It achieves millisecond‑scale latency, high predictive performance, and modest privacy‑utility trade‑offs, offering a practical solution for publishers seeking to balance user privacy with revenue generation.