Modern search engines rely upon extensive mining of users queries and interactions: this is the case also for recent advances in rankers based on pretrained language models. Users rightly worry about the privacy implications of this. For domains such as health, personal and enterprise search, sharing of data is strictly prohibited. In this paper we outline a possible solution for building efective and privacy preserving rankers based on state-of-the-art pretrained language models, and the problems that need to resolved for such a solution to successfully materialise.
and enterprise search, sharing of data is strictly prohibited. Three main barriers are present: 1. Editorial labelling (i.e., labels provided by editors that are not the user/entity who owns the data) would pose a great privacy threat as it requires others to access private data, e.g., an editor reading your WeChat messages. 2. It is unlikely that users would provide a large quantity of explicit labels for their own data, e.g., it does not seem viable to ask you to label a large number of your own Messenger chats. 3. The data itself needs to be transferred to a central search service where the ranker would be trained and operated – and this is often not possible, e.g., consider extracting private medical records from a hospital and allowing a commercial search provider to index and search them. An alternative is to fine-tune a private PLM ranker just on the (likely small) user data and with limited labels: this however will not produce an efective ranker. A key problem then arises: How do we improve a PLM ranker capacity to be efective without large training data, while still preserving the users’ privacy (i.e., without sharing or leaking of their data)?
The intuition that I present in this paper is that a novel framework where PLM rankers are
ifne-tuned in a federated and online manner using implicit user interactions, such as queries
and clicks, may provide solutions for creating efective PLM rankers on private, not shared
data and with no need for explicit labels. The on-device federated learning of a ranker would
allow the data and the ranker to reside directly on the user device while exploiting signals from
multiple users for training the ranker in a privacy-preserving manner, i.e. without sharing the
actual data. Instead of sharing data, clients fine-tune the ranker on their local data and then
communicate to a central server the ranker updates they obtained. Then the server aggregates
the local updates from clients to create a new global ranker, which is then shared back with each
client. This process may be repeated throughout the lifetime of the search product. Exploiting
data from multiple users is vital, as PLM rankers are data-hungry. The online learning process
would allow learning rankers “on-the-fly” via continuous updates and by letting the ranker
probe the search space to gain higher efectiveness and robustness to shifting query intents [
Key to the realisation of the envisioned novel federated online framework for creating efective PLM rankers is the need to address the following challenges:
Challenge 1: How to learn PLM rankers federatively, in an efective and eficient
manner? Our previous work on federated online learning to rank has contributed efective
rankers and aggregation methods [
Challenge 2: What are efective ways to handle noise and bias from clicks when
training PLMs rankers? The envisioned framework relies on learning PLM rankers using
interaction data, such as clicks, in place of editorial labels. PLM rankers in fact require a
substantial amount of data for fine-tuning: the use of readily available interaction data would
make up for the absence of labelled data. Learning from this noisy and biased feedback has been
largely investigated in the context of learning to rank [
Challenge 3: What is the efect of out-of-distribution data, e.g., non identically and independently distributed (non-IID) data? Our previous work on non-IID data in federated online learning to rank has shown the efect of out-of-distribution data on rankers and in which situations this is a problem [15]. This was done for feature-based rankers: whether these problems apply to PLM rankers and to what extent is hard to forecast. We note however that, in non-federated settings, PLM rankers have been already shown to sufer when fine-tuning and testing data are out-of-distribution [16, 17, 18].
Challenge 4: Are federated PLM rankers secure and is user privacy maintained? There are two components that can compromise the security of federated rankers and expose users’ private data. One is the message passing between clients and server; this can generally be secured either via noise injection or via encryption – but their efect on PLM rankers needs investigation. The other component is the actual ranker. Assume a malicious agent joins the federation and accesses the PLM ranker. Can the ranker be “inspected” to gain knowledge about user data and queries? Unlike feature-based rankers, PLMs rankers directly learn from the raw user searchable data (and not its features) and queries. Previous work has shown that PLMs (not for ranking) can expose private information contained in the text they were learned from [19, 20]. How can we then ensure that the model parameters cannot be “reversed” to understand the data used to obtain them?
Note that the use of online learning and of federated learning is not new to Information
Retrieval; in fact, the two have even been already combined in the context of learning to
rank [
I believe that embracing the framework for PLM rankers envisioned in this paper and addressing the challenges outlined above, would allow to create PLM search engines that can be trained in a federated manner, without surrendering user data to a central search service, enabling search vendors to shift ranker creation directly to the user device, a vital requirement when user data should remain not shared.
Acknowledgments This research has been partially sponsored by CCF-Baidu Open Fund. I would like to thank Mr Shengyao Zhuang and Dr Harrisen Scells (IElab, The University of Queensland), and Prof Jimmy Lin (Waterloo University) for comments and suggestions on the ideas exposed here. [13] S. Zhuang, G. Zuccon, Counterfactual online learning to rank, in: ECIR’20, 2020. [14] S. Zhuang, H. Li, G. Zuccon, Implicit feedback for dense passage retrieval: A counterfactual approach, in: SIGIR’22, 2022. [15] S. Wang, G. Zuccon, Is non-iid data a threat in federated online learning to rank?, in:
SIGIR’22, 2022. [16] S. Zhuang, G. Zuccon, Dealing with typos for BERT-based passage retrieval and ranking, in: EMNLP’21, 2021. [17] S. Zhuang, G. Zuccon, Characterbert and self-teaching for improving the robustness of dense retrievers on queries with typos, in: SIGIR’22, 2022. [18] C. Sciavolino, Z. Zhong, J. Lee, D. Chen, Simple entity-centric questions challenge dense retrievers, in: EMNLP’21, 2021. [19] T. Vakili, H. Dalianis, Are Clinical BERT Models Privacy Preserving? The Dificulty of
Extracting Patient-Condition Associations (2021). [20] C. Qu, W. Kong, L. Yang, M. Zhang, M. Bendersky, M. Najork, Natural language understanding with privacy-preserving bert, in: CIKM’21, 2021. [21] E. Kharitonov, Federated online learning to rank with evolution strategies, in: WSDM’19, 2019, pp. 249–257. [22] C. Li, H. Ouyang, Federated unbiased learning to rank, arXiv preprint arXiv:2105.04761 (2021). [23] Q. Yang, Y. Liu, Y. Cheng, Y. Kang, T. Chen, H. Yu, Federated learning, Synthesis Lectures on Artificial Intelligence and Machine Learning 13 (2019) 1–207.