This is the June 2023 edition of AI for Trust, Safety and Fraud, your monthly guide to the latest insights on machine learning for trust, safety, fraud detection, and risk management. This newsletter contains a monthly dose of updates in the field of trust, safety and fraud, and provides useful resources, news and research summaries to stay informed about the constantly evolving field.
In this edition, we will summarize FICO's business model as discussed in an episode of Business Breakdowns podcast, get awareness on the two new MOs of cyber vulnerability, and get an overview on this month's spotlight research where authors combine graph attention and self attention mechanisms to capture inter-variable as well as temporal anomalies in a multivariate time series setting.
(I am always looking for ways to improve the contents of this newsletter, so don't hesitate to share your feedback or suggestions.)
|
|
1. Current Affairs
Operation Triangulation - Security Week
Operation Triangulation is a digital spy campaign that targeted iPhones using iMessages with malicious attachments. The campaign exploited two zero-day vulnerabilities in iOS, one in the kernel and one in WebKit. The vulnerabilities allowed the attackers to execute arbitrary code on the victim's device, which they could then use to install spyware. Ultimately, this vulnerability allowed attackers to get root access to people's iPhone by just delivering them an iMessage, even if users don't click on anything.
The campaign was first reported by Kaspersky Lab in June 2023. Kaspersky researchers said that they had identified the campaign after becoming one of the targets. The researchers said that the campaign was likely targeting Russian government officials and journalists. Apple has since released patches for the two zero-day vulnerabilities. The patches are available for iOS 15.7 and later, as well as for iPadOS 15.7 and later. Operation Triangulation is yet another reminder of the importance of keeping your devices up to date with the latest security patches. If you have an iPhone, make sure that you have installed the latest iOS updates. You can check for updates by going to Settings > General > Software Update. To protect further, you can also disable automatic message preview settings and install an antivirus. |
|
If you suddenly see a live stream of Elon Musk giving bitcoin price predictions on your YouTube recommendations, you might be seeing a new scam happening these days, called Tesla Live Stream Scam. The fraudsters are hijacking YouTube accounts of people with a large following and turned into Tesla or Arc live streams that are designed to scam people. The scammers replay the entire live stream on the hacked accounts and then during parts of the video they will add graphics directing watchers to visit various websites. Those graphics attempt to tell people to send Bitcoin or Ethereum and they will double their coins in 24 hours. Given that these streams are coming from well known YouTubers, many victims tend to believe in the claims and suffer from monetary losses.
|
|
2. FICO - A High Score Business by Business Breakdowns
This section includes a short summary of a discussion by Zack Fuss, an investor at Irenic Capital and Dev Kantesaria, managing partner at Valley Forge Capital Management on FICO's (Fair Isaac Corporation) business model. The discussion is part of a podcast series called Business Breakdowns by Colossus. The notes are modified from originally captured by podcastnotes.org.
FICO is a unique and a successful business that is worth learning more about. I found this episode of Business Breakdowns podcast interesting and thought of adding it into this month’s newsletter. Following are key takeaways.
- FICO earned $1.3 billion in revenue in 2022, out of which 50% is from their credit scoring business (the one we all know about is, car loans, home loans, credit cards etc.) and the rest 50% is from their software offerings, that include fraud protection, customer relationship management, and loan origination.
- 95% of credit scoring revenue comes from North America and is used in over 90% of the credit lending decisions.
- FICO weighted score composition = payment history + debt and credit limits + length of credit history + frequency of new credit + types of credit used
- Newest upgrade to the scoring model is FICO 10 that uses additional data points to make credit scores more accessible and have better predictability.
- Trend data: what has been happening to this consumer over the last 60 days
- Limited history: when you don’t have a substantial credit history, the model can use cable/cell/utility bills to assess credit risk
- Bank accounts: links directly to your bank account to assess activity and balance sheet strength (adds 20 points to your FICO score)
- Resilience index: identifies how resilient consumers are to economic stress compared to other consumers with the same FICO score
- The following factors help in the continued growth of FICO’s business:
- Variety of score uses: The score can be used with different pricing models, in different lending decisions, in different types of organizations, and in both B2b and B2C landscapes. For companies, scores are incredibly cheap compared to the size of lending they need to make.
- Price increases: FICO can pass down price increases without being the bad guy. For example, the credit bureau passes the cost to the financial institution, which then passes it to the consumer in closing costs.
- Lack of competition: The main competitor is Vantage Score but is not really used much in risk assessment decisions. This gives FICO a monopoly-like business model.
- Partnerships: FICO maintains a high quality of their offerings and has a high reputation in the market.
- Software products: FICO smartly uses their years of expertise to build SaaS that help other companies in fraud detection, customer management, and loan origination.
|
|
3. Spotlight Research
Coupled Attention Networks for Multivariate Time Series Anomaly Detection - Feng Xia et al.
This paper by Feng Xia et al., proposes a coupled attention based neural network framework (CAN) to detect anomalies (in unsupervised manner of-course) in multivariate time series setting.
What problem are authors trying to solve?
The authors propose a network architecture to solve Multivariate time series anomaly detection (MTAD). MTAD is the problem that involves finding anomalies in systems that involve multiple data collecting sensors (multiple variables), that capture data in the real time.
What are real world applications of solving this problem?
The real world applications of MTAD include 1) identifying anomalies in industrial control systems, such as power grids, manufacturing plants, and transportation networks, 2) identifying fraudulent transactions in financial services, insurance, and other industries, 3) identifying anomalies in medical data, such as patient vital signs, lab results, and medical imaging, 4) identifying anomalies in environmental data, such as air quality, water quality, and seismic activity. Other important use cases that require automated MTAD include the development and monitoring of smart cities, automated factories, digital twins, public security, self-driving vehicles, and epidemic disease control
What are the issues with existing approaches?
Traditional anomaly detection methods like Isolation Forests can not capture complex inter-sensor (inter-variable) and temporal relationships. GNNs were also already explored to capture inter-variable relationships in MTAD setting, but existing GNN based architectures suffer from lack of prior knowledge about relationships making it harder to define static or dynamic graphs.
How did the authors address it?
The authors actually motivation from "Multivariate Time-series Anomaly Detection via Graph Attention Network (MTAD-GAN) by Hang Zhao et al." (so do read that too), and developed Coupled Attention Network (CAN), on top of it. CAN uses 1) graph learning, 2) attention mechanisms, and 3) convolutional neural networks. They are able to dynamically model both global and local relationships between different variables across time.
Let's open the components of their architecture:
1. Coupled attention module: This module is the primary layer of their architecture. It uses self attention to learn temporal patterns within the same time series, and GCN layer to learn correlations across variables. One key idea of their architecture is "global-local graphs", that basically means a "global graph" represents correlations between entire time series, and "local graphs" further zooming into highly correlated time series to look for local relationships in specific segments of time series.
2. Encoder: The Encoder consists of multiple layers of coupled attention modules to process training samples with position encodings (of-course for any attention layer, positional encodings are required) and encode the inputs to latent vectors that are fed into two decoders as mentioned below.
3. Decoder for prediction: The first decoder takes in historical training samples as well as output of the encoder to create final predictions of anomaly/not anomaly.
4. Decoder for reconstruction: The second decoder is only there to assist in learning representations during training, and not used in inference.
Why did they combine reconstruction and prediction losses in the same architecture?
It is now a common thing in multi-dimensional anomaly detection solutions. The combination of the two can reduce the difficulty of model training, which better characterizes temporal data and captures hidden features between variables.
What experiments were performed to prove that approach works?
The authors conducted experiments on three real-world labeled multivariate time series datasets, with anomalies ranging from 5.85% to 13.13%. They compared multiple methods of anomaly detection including PCA reconstruction error, KNN distance, AE reconstruction error, Isolation Forests, LSTM-VAE, Deep Auto encoding Gaussian Model, OmniAnomaly, MTAD-GAN (very similar to proposed), etc.
They show that their model gives the best F1 score in detecting anomalies on all three datasets.
|
|
Next month, 'AI for Trust, Safety and Fraud' will bring you more thought-provoking articles, industry news, and practical resources. Stay tuned!
The newsletter includes the author's interpretation of news and research works, and can have errors. Please feel free to share feedback, errors or questions, if any.
|
|
|
|
