Today, there are two conflicting views on the nature of data. The first camp believes that data is the “new oil”, as evidenced by a recent Economist article. The position here is that we live in a data driven economy and data is the fuel that powers the new digital economy. The second camp takes a position that data is a “toxic asset”, captured by an article by security expert Bruce Schneier that describes the problems of keeping data and the risk of data getting breached. As with most conflicting views, there is truth to both sides.
Without the right technical capabilities, organizations today are forced to to make a hard trade-off in picking one of these two camps. Either they ignore the risk of data sharing and data retention and end up erring on the side of extracting value out of the data for organizational workflow or some organizations err on the side of caution with this risk and organizational workflows might end up being stalled.
In this broader context, it is worth revisiting data sharing through the lens of the principle of least privilege from the discipline of computer security. The classic principle of least privilege dictates that users, applications, and systems should be granted only the minimum necessary access rights to perform their tasks. This concept aims to reduce the attack surface and limit the potential damage from security breaches by minimizing the privileges assigned to each entity.
Let us consider a simple hypothetical example where Alice and Bob want to collaborate. Alice is the data owner or data holder and Bob is the consumer of the data who needs access to Alice’s data set. In this case, Bob has a task they need to do using the data and is blocked without access to the data. Suppose the data set is the list of all employees salaries in an organization and Bob wants to compute the average salary as the task. From an information theoretic perspective in this case the only information Bob needs is the average of the salary and does not need access to other fields or individual salary information. Thus, applying the principle of least privilege, the “minimal” information content Alice needs to share with Bob to unlock Bob would just be the average or a single number in this case; a sufficient set of statistics for the task.
In this example, Bob’s task was simple and very well defined in terms of a closed form, simple arithmetic function that Alice could just calculate and reveal the answer. But what happens when Bob’s task is far more complicated or cannot be expressed using such a close form function or Bob does not know ahead of time what kind of exploratory data analysis he needs to do for his task. Maybe Bob does not know that he only wants the average but wants both the average and the median. Or Bob wants to slice and dice this data set by different departments and get the average per department. Or Bob really wants to build a machine learning model to check if there are intrinsic biases in the salary? In such more complex scenarios, the data sharing problem becomes a lot more nuanced as it may be challenging to precompute sufficient statistics. This gets worse with high dimensional data with multiple fields and measurements.
Now, Alice and Bob have a conundrum. Alice wants to collaborate with Bob to help Bob’s workflow from being blocked but at the same time is wary of sharing the entire data set in the raw form. Alice could try to pre-compute a variety of queries and statistics and share them with Bob, but as we saw that might be insufficient for Bob’s possibly unforeseen and undeclared future tasks. Alice could consider anonymizing the data and sharing with Bob, but maybe worried that anonymization is insufficient to meet her organization’s sharing policies. In particular, in particular, anonymization retains a one-to-one correspondence between every entry in the source data and the shared data, which could become problematic. Alice and Bob could also consider more heavyweight machinery like federated learning (in the specific case the task is a predefined machine learning algorithm model to be learned) or computing on encrypted data, but these might require significant effort and modifications to existing workflows, and may not be compatible with the tasks that Bob wants to achieve as seen in the qualitative comparison above.
Recent advances in Generative AI offer an elegant solution to Alice and Bob’s conundrum! Instead of sharing just a handful of hardcoded summary statistics or sharing the raw or anonymized data, Alice could use modern Generative AI techniques to create a generative model of her data set and then use this generator model to create privacy, preserving synthetic data set to share with Bob. In essence, Alice is leveraging the ability of deep generative models to act as high dimensional function, approximates that capture the essential statistical and emergent properties of a data set and use that as a basis for sharing data set safely with Bob. Said differently, Alice is creating a “deep fake” of her original data set that she can safely share with Bob which preserves the essential structure that Bob needs for his task without unnecessarily revealing other aspects of the dataset that the organizational policies need to protect.
In effect, synthetic data via deep generative models offer an elegant technical basis for a solution to this long-standing conundrum of data sharing for collaborative workflows both within and between organizations. We can view synthetic data as offering a technological basis for balancing this trade-off between least privilege and the essential information needed for the downstream task. By carefully orchestrating synthetic data based workflows, organizations can now have the best of both worlds by ensuring workflows are not blocked by data access while still only sharing a minimum amount of information in the restricted datasets.
Authored by Vyas Sekar, Co-founder & Chief Technologist at Rockfish.Connect on LinkedIn.