TLDR: This research paper investigates fraud on subscription platforms, where existing machine learning detection methods are in an ‘arms race’ with bad actors. It critiques the widely used GLOBALPROP revenue division rule, showing it fails to prevent fraud and makes detection computationally intractable. The paper introduces three manipulation-resistance axioms (fraud-proofness, bribery-proofness, Sybil-proofness) and proposes a novel rule, SCALEDUSERPROP. This new mechanism satisfies all three axioms and is empirically shown to be a fairer alternative to existing rules like USERPROP and USEREQ, offering an inherent disincentive against fraudulent activities.
Subscription-based platforms, from music streaming giants like Spotify and Apple Music to video services, have become a cornerstone of digital content consumption. Users pay a fixed fee for unlimited access, and content creators receive a share of the revenue. However, this model is increasingly vulnerable to sophisticated fraud, costing the industry hundreds of millions annually. A recent case highlighted this issue when a musician was criminally charged for orchestrating a scheme that fraudulently inflated his music streams, earning over US$10 million in royalties.
Traditionally, platforms combat this fraud using machine learning (ML) methods. While effective to a degree, this approach often leads to an ongoing “arms race” with bad actors who continuously evolve their manipulation tactics, often leveraging advanced AI-generated content and bot networks. This research paper, titled Fraud-Proof Revenue Division on Subscription Platforms, explores a different path: designing revenue division mechanisms that inherently discourage manipulation, rather than relying solely on detection after the fact.
The Problem with Current Revenue Sharing
The core of the problem lies in how revenue is typically distributed. Many major streaming platforms use a rule called GLOBALPROP. Under GLOBALPROP, funds from the total royalty pool are allocated to artists proportionally based on their share of total streams across the entire platform. This means that a user who streams music extensively can have a disproportionate influence on revenue distribution. This characteristic makes GLOBALPROP highly susceptible to fraud, as bad actors can create fake users (bots) to artificially boost streams for their content, turning a profit if the royalties generated exceed the cost of maintaining the fake subscriptions.
The researchers demonstrate that GLOBALPROP not only fails to prevent fraud but also makes detecting suspicious activity computationally intractable, meaning it’s incredibly difficult for platforms to identify who is benefiting most from fraudulent activities. This is a significant finding for the ML community, which often focuses on detection.
Designing for Manipulation Resistance
The paper introduces three key axioms to formalize manipulation resistance:
- Fraud-proofness: Prevents adversaries from profitably creating new fake users.
- Bribery-proofness: Prevents artists from profitably bribing existing users to manipulate their engagement profiles.
- Sybil-proofness: Ensures artists cannot gain an unfair advantage by splitting into multiple identities or merging with others.
The study found that while GLOBALPROP fails to satisfy fraud-proofness and bribery-proofness, two other existing rules, USERPROP and USEREQ, do satisfy these manipulation-resistance axioms. USERPROP distributes each user’s subscription fee only among the creators that user consumes, proportional to their engagement. USEREQ distributes each user’s fee equally among the artists they engage with. However, these rules have their own trade-offs in terms of fairness and other properties.
Introducing SCALEDUSERPROP: A Fairer Alternative
To address the limitations of existing rules, the researchers introduce a novel rule called SCALEDUSERPROP. This mechanism aims to strike a balance between manipulation resistance and fairness. SCALEDUSERPROP works by having the platform take a disproportionate amount of commission from low-engagement users. It then distributes the remaining subscription fees using a method similar to USERPROP.
Essentially, SCALEDUSERPROP can be seen as a variant of GLOBALPROP that “limits the influence” of users who have engagement significantly above average. This design ensures that it satisfies all three manipulation-resistance axioms (fraud-proofness, bribery-proofness, and Sybil-proofness), similar to USERPROP. Crucially, experiments conducted with both real-world and synthetic streaming data support SCALEDUSERPROP as a fairer alternative compared to existing rules, especially for varying revenue share percentages (alpha values).
Also Read:
- Securing Decentralized Data Selection: A New Approach to Robust Federated Submodular Maximization
- Enhancing AI Safety: New Research on Robust Control Protocols
Looking Ahead
This research provides a principled framework for designing revenue division mechanisms that are inherently resistant to fraud. By shifting the focus from reactive fraud detection to proactive mechanism design, platforms can build a more robust and equitable ecosystem for content creators. The findings suggest that SCALEDUSERPROP offers a promising path forward, providing strong manipulation resistance while also promoting fairer revenue distribution.
