10 min reading
In model training in Artificial Intelligence, the size of the data set with enough variance is defining in many cases the accuracy of the model for the use in the wild. With a limited set of data available for model training, a common technique is to create useful complementary data that augment the original data, called synthetic data.
This idea applied from the other side of trustworthy AI can lead to serious problems in the society.
Time is money
The idea of training data and test data is simple. With enough test data, based on an existing, pre-trained model, one can simply figure out how well the model is working. A common challenge is to provide this “enough” of test data. To work around this challenge, data can be also created to be used as test data if not enough real world data is available.
The timeline is usually an issue, too: having a model developed for a certain use case, it’s time to release it for production. Before going to release it, a massive set of test data is always welcomed but if the desired use case is not that common throughout the customer base yet – because it’s still being developed – real world data is rare and usually only available from lab testing, ideally performed at scale.
For that purpose, creating a set of synthetic data might ease the pain of being unsure about the accuracy of the model. Synthetic data, differently from real data, can also include specific deviations that may happen, let it be in the changed number of occurrences, the range of a parameter, etc. This could help in testing the model for it’s behavior under changes to the input data and might test for the representation of a certain kind of drift. Knowing this might help to assess drift in production and to quickly identify reasons and changed patterns that influenced the outcome.
From the perspective of the model, there is no difference between real world and synthetic data. For a data scientist, the richness of the synthetic data with all the outliers or additional parameters ignored for the model, based on different reasons, might indicate a difference when comparing those sets for their origin. Still, for the real world data used later on with the model in production and thus for the real world applying the model for any purpose, there will be no difference.
A good practice
The challenge for the real world instead is starting when the synthetic data is provided as real world data. … And this might be not about the model applied for production.
Bringing a model into production for a certain use case can provide better understanding of the model. Initially, only a few customers will use it and the model viability is tested against this small but additional set of real world data.
With an established customer base and enough different situations the use case is applied, a large enough set of test data becomes available to test for necessary changes of the model, implicit because of monitoring or explicit when customers provide the data in addition.
To bridge the initial gap between this desired state and the state before release of the model to a target infrastructure (use case), creativity is needed. It’s especially needed to create artificial test data that should look like real world data, ideally in full variance, breath, and depth, like a large set of diverse environments would provide. Since the use case is known and the data format as well, as well as future versions of the part producing the data for consumption in the use case like a set of well defined sensors, creating data would be easier than creating a massive amount of infrastructures producing the data in the real world. This is common practice throughout the industry and there is nothing bad about it.
Under attack
As with the discipline of data engineering for creating synthetic data is expanding, the border line between real and artificial data might become blurred when the use of those data gets shifted towards unethical cases. Or, simply used for cheating.
Not in general but… just in case that one wouldn’t have enough data for providing evidence on a specific subject for – let’s say, science in general – what to do ? Many scientific journeys for academic research projects usually come under pressure if no time is left to achieve meaningful outcomes to justify an investment made for investigating a certain aspect in the real world. And since a failed investment perhaps doesn’t lead to a new sponsoring with a new investment, the results might not be provided and credibility and careers might be on risk. Then, creativity at its worst might be a solution: fake the data.
We are not good in reviewing massive data or just only complex enough data, but Artificial Intelligence is good in this regard if properly trained. This efficiency on complex tasks is good on assessing data. The very same tool, however, can be creatively used to also produce data. So, why not (ab)using the data science for creating synthetic data for a kind of data set that would help to underpin any hypothesis ? Human generated data might be identified at some point because the stochastic abnormalities will be missed or will reveal some kind of a pattern not expected. With the help of AI, perhaps well crafted data can be generated with enough tuning towards masquerading the usual human pitfalls. And with enough data generated, a human will barely check the viability without using tools.
The challenge for this task is to trick another AI model that might be used to check the data integrity and viability of the data sets provided. It seems that spending the time on data engineering for some projects might be a more achievable way of using AI than collecting and assessing real world data. And there are two main reasons: it’s a determined way to achieve a result good enough to showcase the approach stated initially, and it’s more cost effective than running around and collecting data.
Especially, with all the environmental data to collect over time for environmental research, or for human behavioral studies, the effort might be significantly lower than preparing for an investigation in the real world and finding the right spots for collecting appropriate data, getting the right (matching) people over a prolonged time, or even getting access to locations for the desired data collection, etc.
With the high pressure of showing success to stakeholders, cheating becomes a viable option because nobody would be able to check all the potentially real data or fake data for their viability and validity. And the more people would addict to this inethical approach of doing science the more it would become impossible to check and validate the massive amount of provided “real world data” sets.
Most of scientific investigations rely on sponsoring or orders. It’s difficult to be supported in any of those kinds with no previous reliable findings. It’s important to secure resources, being it first and foremost time, people and material resources, slots for using specific equipment. However, by creating data, it’s becoming easy then to act as a scientist in a specific direction or supporting directions of research with data that might be of special interest to various groups in our society.
Applying such a (theoretical ?) misbehavior is dangerous by itself. It’s simply not only cheating but results in pure manipulation where science becomes obscure. It provides a strange opportunity to betray the society by obscuring the most important tool of checking with the real world and our models about it we base our daily life and future on.