TL;DR
This paper introduces a novel attack scheme that exploits the common "gold standard" method for quality control in crowdsourcing. By colluding, workers can build a system to identify and avoid these gold questions, allowing them to achieve high accuracy while significantly speeding up their work and increasing their earnings.
What Problem Does This Paper Address?
Crowdsourcing platforms are widely used to collect large-scale datasets for training AI models. A critical challenge in this process is ensuring the quality of the data provided by human annotators. One of the most popular quality control mechanisms is the use of "gold questions" – tasks with known correct answers that are interspersed with regular tasks. These gold questions are used to assess annotator accuracy in real-time and flag low-quality contributors. However, the effectiveness of this method relies on the assumption that workers cannot easily identify which questions are gold. This paper addresses the vulnerability of this gold standard approach, demonstrating how it can be undermined by a coordinated attack.
Methods
The core of the attack involves a group of colluding crowd workers who deploy a decentralized system. This system consists of:
- Browser Plugin: This plugin runs on the worker's browser and analyzes the HTML of the crowdsourcing page. It anonymizes worker information, removes identifying tags, splits tasks, and then generates a "fingerprint" for each task using simhashing. Simhashing is a technique that creates compact representations of documents, allowing for efficient similarity detection.
- External Server: The fingerprints generated by the plugins are sent to a central server. This server maintains a repository of fingerprints and their "multiplicity" (how many times each fingerprint has been seen).
- Inferential System: The server uses the multiplicity data to infer which tasks are likely to be gold questions. The key assumption here is that gold questions, being a smaller subset, will appear more frequently among the colluding workers' fingerprints than non-gold questions. A Gaussian Mixture Model is used to identify these bimodal distributions and classify tasks.
The attack is designed to be anonymous (workers don't need to be identified) and secure (information doesn't need to be shared directly between all workers).
Datasets
The experimental evaluation of the proposed attack scheme was performed on the CSTA dataset. This dataset contains 29,594 judgments from 336 workers, including timestamps and worker answers. The paper notes that the original dataset has a gold question ratio of 12.4%, and for their experiments, they subsampled this set to simulate smaller gold question ratios (e.g., 4.4%).
Key Insights and Ideas
- Gold Questions are Predictable: The paper demonstrates that even with a relatively small proportion of gold questions, their frequency can be exploited to identify them.
- Decentralized Attack is Effective: The combination of a browser plugin and an external server allows for a coordinated attack without direct communication between all colluding workers.
- Speed vs. Accuracy Trade-off: Attackers can achieve high accuracy on gold questions (thus avoiding detection) while significantly reducing the time spent on non-gold questions, leading to a substantial increase in hourly wage.
- Countermeasures Have Limitations: The paper explores several potential countermeasures, such as increasing gold set size or varying question selection, but finds them to be either costly or not significantly effective against the attack.
Key Results
- High Accuracy Maintained: The proposed attack method allows colluding workers to maintain high accuracy on gold questions, making them appear as legitimate workers.
- Significant Time Savings: Workers using the attack scheme can save a considerable amount of time, leading to hourly wages up to five times higher than the baseline approach. For instance, with a 4.4% gold ratio, the time saved per page was consistently high.
- Robustness to Gold Ratio: The attack remains effective even with higher gold question ratios (e.g., 12.4%), although it may take longer for the inferential system to reach high detection accuracy.
- Countermeasures Ineffective: Experiments showed that varying the selection of gold questions (e.g., serving the least seen question) did not significantly improve accuracy compared to uniform random selection, especially with small gold sets.
Limitations and Future Work
- Assumption on Gold Set Size: The attack relies on the assumption that the gold set is significantly smaller than the set of non-gold questions. If gold questions were as numerous as regular tasks, the attack would likely fail.
- Assumption on Sampling: The initial assumptions about uniform random sampling of gold questions and no repetition for each worker are relaxed later, but the core attack still holds.
- Future Work: The authors suggest exploring extensions where colluding workers share additional information, such as answers to gold questions, which could further increase efficiency benefits. They also mention that programmatic gold questions, while potentially helpful, increase setup and design costs.
Summary updated