Jan 26, 2021

IBM develops software to reduce personal data in AI training

William Smith
2 min
The IBM AI Privacy and Compliance Toolkit allows data scientists to create machine learning models that protect the privacy of training data
The IBM AI Privacy and Compliance Toolkit allows data scientists to create machine learning models that protect the privacy of training data...

Researchers at US technology giant IBM have developed ways of improving the protection of privacy during the training of artificial intelligence models.

The AI Privacy and Compliance Toolkit allows data scientists to create machine learning models that protect the privacy of training data while following the necessary data protection regulations.

Overcoming AI security issues

The issue is that, even if training data itself is not exposed, AI trained on real data might leak sensitive information if someone is determined enough.

The IBM software, which assesses privacy risk, has applications in industries ranging from fintech to health care to insurance - anywhere that relies on sensitivity training data. The software involves a number of approaches, including differential privacy (DP).

In a blog post, Abigail Goldsteen, Researcher in Data Security & Privacy, IBM Research, said: “Applied during the training process, DP could limit the effect of anyone’s data on the model’s output. It gives robust, mathematical privacy guarantees against potential attacks on a user, while still delivering accurate population statistics. [...] However, DP excels only when there’s just one or a few models to train. That’s because it’s necessary to apply a different method for each specific model type and architecture, making this tool tricky to use in large organizations with a lot of different models.”

The specifics

To counteract that, data can be anonymised before the model is trained. The process involves generalising data, by removing specific values and instead providing a blurred range. IBM’s innovation in its software is to tailor the extent of that process to the needs of the organisation.

“This technology anonymizes machine learning models while being guided by the model itself,” said Goldsteen. “We customize the data generalizations, optimizing them for the model’s specific analysis – resulting in an anonymized model with higher accuracy. The method is agnostic to the specific learning algorithm and can be easily applied to any machine learning model.

Share article

Jun 17, 2021

Facebook Develops AI to Crackdown on Deepfakes

3 min
Social media giant, Facebook, has developed artificial intelligence that can supposedly identify and reverse-engineer deepfake images

In light of the large tidal wave of increasingly believable deepfake images and videos that have been hitting the feeds of every major social media and news outlet in recent years, global organisations have started to consider the risk factor behind them. While the majority of deepfakes are created purely for amusement, their increasing sophistication is leading to a very simple question: What happens when a deepfake is produced not for amusement, but for malicious intent on a grander scale? 


Yesterday, Facebook revealed that it was also concerned by that very question and that it had decided to take a stand against deepfakes. In partnership with Michigan State University, the social media giant presented “a research method of detecting and attributing deepfakes that relies on reverse engineering from a single AI-generated image to the generative model used to produce it.” 


The promise is that Facebook’s method will facilitate deepfake detection and tracing in real-world settings, where the deepfake image itself is often the only information detectors have to work with. 

Why Reverse Engineering? 

Right now, researchers identify deepfakes through two primary methods: detection, which distinguishes between real and deepfake images, and image attribution, which identifies whether the image was generated using one of the AI’s training models. But generative photo techniques have advanced in scale and sophistication over the past few years, and the old strategies are no longer sufficient. 


First, there are only so many images presented in AI training. If the deepfake was generated by an unknown, alternative model, even artificial intelligence won’t be able to spot it—at least, until now. Reverse engineering, common practice in machine learning (ML), can uncover unique patterns left by the generating model, regardless of whether it was included in the AI’s training set. This helps discover coordinated deepfake attacks or other instances in which multiple deepfakes come from the same source. 


How It Works 

Before we could use deep learning to generate images, criminals and other ill-intentioned actors had a limited amount of options. Cameras only had so many tools at their disposal, and most researchers could easily identify certain makes and models. But deep learning has ushered in an age of endless options, and as a result, it’s grown increasingly difficult to identify deepfakes.


To counteract this, Facebook ran deepfakes through a fingerprint estimation network (FEN) to estimate some of their details. Fingerprints are essentially patterns left on an image due to manufacturing imperfections, and they help identify where the image came from. By evaluating the fingerprint magnitude, repetition frequency, and symmetrical frequency, Facebook then applied those constraints to predict the model’s hyperparameters. 


What are hyperparameters? If you imagine a generative model as a car, hyperparameters are similar to the engine components: certain properties that distinguish your fancy automobile from others on the market. ‘Our reverse engineering technique is somewhat like recognising [the engine] components of a car based on how it sounds’, Facebook explained, ‘even if this is a new car we’ve never heard of before’. 


What Did They Find? 

‘On standard benchmarks, we get state-of-the-art results’, said Facebook research lead Tal Hassner. Facebook added that the fingerprint estimation network (FEN) method can be used for not only model parsing, but detection and image attribution. While this research is the first of its kind, making it difficult to assess the results, the future looks promising. 

Facebook’s AI will introduce model parsing for real-world applications, increasing our understanding of deepfake detection. As cybersecurity attacks proliferate, and generative AI falls into the hands of those who would do us harm, this method could help the ‘good guys’ stay one step ahead. As Hassner explained: ‘This is a cat-and-mouse game, and it continues to be a cat-and-mouse game’.

Share article