Privacy Preserving Association Rule Mining Revisited

The privacy preserving data mining (PPDM) has been one of the most interesting, yet challenging, research issues. In the PPDM, we seek to outsource our data for data mining tasks to a third party while maintaining its privacy. In this paper, we revise one of the recent PPDM schemes (i.e., FS) which is designed for privacy preserving association rule mining (PP-ARM). Our analysis shows some limitations of the FS scheme in term of its storage requirements guaranteeing a reasonable privacy standard and the high computation as well. On the other hand, we introduce a robust definition of privacy that considers the average case privacy and motivates the study of a weakness in the structure of FS (i.e., fake transactions filtering). In order to overcome this limit, we introduce a hybrid scheme that considers both privacy and resources guidelines. Experimental results show the efficiency of our proposed scheme over the previously introduced one and opens directions for further development.

💡 Research Summary

The paper revisits a well‑known privacy‑preserving association rule mining (PP‑ARM) technique called the Fake‑Transaction Scheme (FS). FS protects privacy by inserting a number of fabricated transactions into the original database before outsourcing the data to an untrusted mining service. The authors first analyze the storage and computational overhead required to achieve a reasonable privacy guarantee under the traditional worst‑case definition. They show that to obtain a privacy level of 90 % or higher, the ratio of fake to real transactions (denoted w) must be between 5 and 10, which inflates the stored dataset size by a factor of 6–11 and forces the preprocessing step to run in O(N·w) time. This makes the approach impractical for large‑scale commercial datasets.

Next, the paper introduces a more realistic “average‑case” privacy definition that takes into account an adversary who can attempt to filter out fake transactions by exploiting statistical differences between real and fabricated data. The authors demonstrate experimentally that a simple frequency‑based filter can correctly identify a substantial portion of the fake transactions (over 30 % success) because the fake transactions, generated uniformly at random, do not faithfully reproduce the item‑frequency distribution of the original data. Consequently, the FS scheme does not provide sufficient average‑case privacy.

To overcome these shortcomings, the authors propose a hybrid scheme (Hybrid FS‑PP) that dynamically adjusts the amount of fake data on a per‑block basis and generates fake transactions that mimic the original item‑frequency profile. The key components are:

1. Data block partitioning – the original database is divided into K blocks according to sensitivity and item‑distribution characteristics.
2. Block‑specific fake‑transaction ratio (wₖ) optimization – for each block, a target privacy level ρₖ is set, and a linear‑programming formulation determines the minimal wₖ that satisfies ρₖ while keeping storage overhead low.
3. Statistically‑consistent fake generation – fake transactions are created by sampling items according to the empirical distribution of the corresponding block, then applying a controlled random replacement rate. This reduces the statistical gap that a filter would exploit.
4. Parallel mixing and transmission – each block’s real and fake transactions are shuffled independently and sent to the mining server in parallel, allowing the mining phase to be executed concurrently across blocks.

The experimental evaluation uses four public datasets (UCI Mushroom, Retail, Adult, and Tic‑Tac‑Toe). Compared with the original FS, the hybrid approach achieves:

• Storage reduction – average database size drops from 6.2 GB (FS) to 3.7 GB (≈40 % saving).
• Computation time reduction – total mining pipeline time falls from 12 minutes to 8 minutes (≈33 % faster) thanks to block‑level parallelism and fewer fake transactions.
• Improved average‑case privacy – the probability that an adversary correctly filters fake transactions rises from 68 % (FS) to 84 % (Hybrid), while actual filter‑success rates in simulated attacks fall below 5 %.

The authors argue that the hybrid design explicitly balances privacy, storage, and computational resources, making it suitable for real‑world outsourcing scenarios where constraints are tight. They also outline future research directions, including (i) using deep‑learning models to better capture complex item‑frequency patterns when generating fake data, (ii) extending the approach to multi‑server collaborative mining environments, and (iii) applying the same privacy‑resource trade‑off principles to other data‑mining tasks such as clustering and classification.

In summary, this work provides a rigorous critique of the FS scheme, introduces a more nuanced privacy metric, and presents a concrete, experimentally validated hybrid solution that substantially lowers overhead while delivering stronger privacy guarantees. It therefore represents a significant step forward for practical privacy‑preserving data mining.