TL;DR
This paper introduces "active honeytokens" – fake data objects designed to detect insider data misuse. They found that users are surprisingly good at distinguishing these tokens, but knowing about honeytokens doesn't change their behavior, and a higher density of honeytokens significantly improves detection rates.
What Problem Does This Paper Address?
Organizations rely on sensitive data, and its misuse can be devastating. While traditional security focuses on external threats, insider threats (employees misusing their access) are a significant concern. Detecting this misuse is challenging because insiders often have legitimate access and can blend in. The paper addresses how to effectively detect insider data misuse by using "active honeytokens" – a type of honeypot embedded within real data. The core questions are: can users distinguish these fake tokens from real data, and does knowing about them change their behavior or the effectiveness of detection?
Methods
The authors developed a system called HoneyGen for automatically generating "good" honeytokens. This system works in three phases:
- Rule Mining: Extracts rules describing the structure, characteristics, and constraints of real data.
- Honeytoken Generation: Creates artificial data objects (honeytokens) that satisfy these extracted rules, ensuring they are consistent with the real data.
- Likelihood Rating: Assigns a "likelihood score" to each honeytoken, indicating how similar it is to real data. Higher scores mean the honeytoken is more indistinguishable.
They then conducted two experiments:
- Experiment I: A Turing-like test where participants tried to distinguish real data from honeytokens with varying likelihood scores.
- Experiment II: A behavioral study simulating a loan approval process where participants acted as bankers. This experiment tested the impact of awareness of honeytokens, the density of honeytokens, and the value of the data object on misuse behavior and detection rates.
Datasets
- Experiment I: Used profiles from a dating website. They collected 1,093 real profiles and generated 13,060 honey-profiles using HoneyGen.
- Experiment II: Used the Statlog (German Credit Data) dataset from the UCI repository. They selected 11 features from this dataset and generated honey-loan requests.
Key Insights and Ideas
- "Good" Honeytokens are Hard to Spot: Honeytokens generated by HoneyGen with high likelihood scores are very difficult for humans to distinguish from real data. This is crucial for their effectiveness.
- Awareness Doesn't Deter Misuse: Surprisingly, knowing that honeytokens are present doesn't significantly change users' tendency to misuse data. This suggests that detection mechanisms need to be robust, as users won't necessarily self-censor just because they know they're being watched.
- Density is Key: Increasing the proportion of honeytokens in a dataset significantly boosts the detection rate of malicious activity. More fake data means a higher chance of a misuser stumbling upon it.
- Value Matters: In the loan approval experiment, participants were more inclined to approve "attractive" (high-value) loan requests illegally, highlighting that the perceived profit strongly influences misuse behavior.
- Religion as a Deterrent: Interestingly, religiosity was found to be a statistically significant factor in reducing illegal actions, aligning with previous research on honesty.
Key Results
- Experiment I: Honeytokens with likelihood scores higher than the median score of real tokens were identified correctly only about 50% of the time, indicating they were indistinguishable.
- Experiment II:
- Participants illegally approved about 32% of loan requests on average.
- When honeytokens constituted 20% of the list, the detection rate was 100% for both informed and uninformed groups.
- A higher percentage of honeytokens (20% vs. 10%) led to a statistically significant increase in detection rates.
- The logistic regression showed that %Honeytokens, LoanAmount, Age, and Religiosity were statistically significant factors influencing illegal approval.
Limitations and Future Work
- Dataset Distribution: The honey-loan requests in Experiment II were generated with a uniform distribution, not reflecting the actual distribution of loan values in the real dataset. This might have influenced the results, as higher average loan amounts in honeytokens made them more attractive. Future work should incorporate more realistic data distributions into HoneyGen.
- Trade-off Between Detection and Cost: The paper acknowledges the need to evaluate the trade-off between the effectiveness of honeytokens and their inherent costs (creation, maintenance, monitoring).
- Ecological Validity: The experiments used datasets (dating profiles, loan requests) and participants (students) that may not perfectly mirror real-world scenarios. A field trial with actual organizations and employees is suggested for better validation.
- Consequences: The study participants faced minimal consequences for illegal actions (loss of bonus points), unlike real-world scenarios where termination or legal action is possible.
- Reward Influence: The authors suggest exploring how manipulating rewards could influence user behavior, which is an interesting avenue for future research.
Summary updated