Artificial intelligence across company borders

Olga Fink, Torbjørn H. Netland, Stefan Feuerriegel

TL;DR

This paper proposes a framework for "cross-company AI" by combining Federated Learning with Domain Adaptation. By allowing companies to train collaborative models without sharing raw data, the authors aim to unlock the $13 trillion potential of industrial AI while navigating the hurdles of data privacy and heterogeneous operating conditions.

What Problem Does This Paper Address?

Artificial Intelligence thrives on data, but in the industrial sector, the most valuable data—representing rare system failures or critical outages—is often scarce for any single company. While the obvious solution is for companies to pool their data, two major hurdles stand in the way:

  1. Privacy and Intellectual Property: Companies are rightfully reluctant to share raw data that might reveal trade secrets, operational throughput, or proprietary parameter settings to competitors or customers.
  2. Domain Shift: Even if data is shared, a model trained on Company A’s machinery might fail at Company B because of different configurations, environments, or operating conditions. This discrepancy in data distribution is known as "domain shift."

The authors argue that these barriers have left a massive amount of economic potential untapped, particularly in safety-critical sectors like power systems and manufacturing.

Methods

The paper advocates for a hybrid approach called Federated Transfer Learning, which merges two distinct AI paradigms:

  • Federated Learning (FL): Instead of moving data to a central server, the model is moved to the data. Companies train local models on their own servers and only share model parameters (like gradients or weights) with a central aggregator or each other. This ensures raw data never leaves company borders.
  • Domain Adaptation (DA): To solve the "domain shift" problem, the authors suggest using DA techniques to learn "invariant representations." The goal is to extract features that are useful for the task (e.g., predicting a fault) but indistinguishable between different companies' datasets.

The authors also discuss the evolution from centralized federated learning (which relies on a central coordinator) to decentralized federated learning, potentially utilizing Distributed Ledger Technology (DLT) to enable peer-to-peer collaboration without a single point of failure.

Datasets

Not discussed in the paper. As this is a "Viewpoint" article rather than an empirical study, the authors do not perform experiments on specific datasets. Instead, they provide a conceptual framework and cite industrial examples such as power system outages and machine fault prediction as primary use cases.

Key Insights and Ideas

  • Vertical Collaboration: Unlike traditional FL, which often focuses on millions of mobile devices (horizontal), cross-company AI is "cross-silo" or "feature-based," involving a smaller number of entities with highly specialized data.
  • The Necessity of Invariance: The authors highlight that FL alone is insufficient for industry. Without Domain Adaptation, a collaborative model would likely be biased toward the company with the most data or the most "standard" operating conditions, making it useless for others.
  • AI-as-a-Service: The framework opens the door for new business models where third-party orchestrators provide the federated infrastructure, allowing small and medium-sized enterprises (SMEs) to benefit from "population-wide" insights they could never generate alone.
  • Systems Thinking: The paper emphasizes that this isn't just a technical challenge but a "service systems engineering" one, requiring a systematic understanding of how value is co-created across different stakeholders.

Key Results

Because this paper is a conceptual framework and position piece, it does not provide quantitative results (such as accuracy metrics or speedup benchmarks). The primary "result" is the proposed architectural roadmap for implementing privacy-preserving, domain-aware AI in industrial ecosystems.

Limitations and Future Work

The authors are transparent about several remaining challenges:

  • Privacy Guarantees: Standard Federated Learning does not offer a 100% guarantee of privacy; it is sometimes possible to reverse-engineer information from gradient updates. The authors suggest incorporating differential privacy or cryptography as additional layers.
  • Security Vulnerabilities: The system could be susceptible to "causative attacks," where a malicious participant provides false model updates to corrupt the collaborative model.
  • Economic Incentives: There is a need for "fair value-capture distribution models." If Company A provides better data than Company B, how is that value quantified and rewarded?
  • Standardization: For this to work at scale, the industry needs consistent standards for data heterogeneity and continuous learning to handle systems that change over time.

Summary updated

Abstract

Artificial intelligence (AI) has become a valued technology in many companies. At the same time, a substantial potential for utilizing AI \emph{across} company borders has remained largely untapped. An inhibiting factor concerns disclosure of data to external parties, which raises legitimate concerns about intellectual property rights, privacy issues, and cybersecurity risks. Combining federated learning with domain adaptation can provide a solution to this problem by enabling effective cross-company AI without data disclosure. In this Viewpoint, we discuss the use, value, and implications of this approach in a cross-company setting.
BibTeX
@article{fink2022artificial,
  author = {Olga Fink and Torbjørn H. Netland and Stefan Feuerriegel},
  title = {{Artificial intelligence across company borders}},
  journal = {Communications of the ACM},
  year = {2022},
  doi = {10.1145/3470449},
}

Loading related content...

Figure