Traditional anomaly detection methods are either rule-based, which doesn’t generalize well since the rules are too specific to cover all possible scenarios or time-series based, (time vs. quantity) which is too low-dimensional to capture the complexity of real life. Real life events have higher dimensions (time, both source and destination locations, activity-type, object-acted on, app used, etc.) A successful anomaly detection system will have eight “must-have” features.
Before we go through those features, at the highest level the system must be one that “whitelists” rather than “blacklists” and is based on machine learning.
The reason why whitelisting is critical is because it studies the good guys. Bad guys try to hide and outsmart blacklist-based platforms like anti-malware. A successful machine-learning anomaly detection system won’t chase bad guys, looking for “bad-X” in order to react with “anti-X.” Instead, such a platform that is whitelist-based can study what is stable (good guys’ normal behavior) and then look out for outliers. This approach avoids engaging in a perpetual and futile arms race.
If you’re going to do anomaly detection the right way, you need to be able to scale to billions of events per day and beyond. It’s not practical at that scale to define whitelists a-priori, or keep a perfect history of all observed behavior combinations. Instead, anomaly detection models should be “soft” in the sense that they always deal with conditional probabilities of event features and are ever-evolving.
The second high-level requirement is that a successful anomaly detection system must be machine learning-based. Virtually every CASB today uses this term, but few mean it. Machine learning means just what it says, that pattern recognition should be done by the computer without being specifically told what to look for. There are two main types of machine learning: Supervised and unsupervised. The former is where the computer learns from a dataset of labeled training data whereas the latter is where the computer makes sense of unlabeled data and finds patterns that are hard to find otherwise. Both supervised and unsupervised machine learning are relevant for this blog, and from here on out I’ll simply refer to anomaly detection as “Machine Learned Anomaly Detection,” or “MLAD” for short.
Now that we have established some high-level requirements, let’s dive into the eight “must-haves” for effective MLAD.
Noise resistance: A common issue with all anomaly detection systems is false-positives. In reality, it’s hard to avoid false positives entirely because in the real world there’s always an overlap between two distributions with unbounded ranges and different means. The chart below, which includes two distributions from the same data set of test results, shows this. Move the criterion threshold value to the right and you get fewer false-positives (FPs). The problem is that by doing this you’ll be also getting a growing number of false negatives (FNs). There is always a tradeoff.
While it is difficult to avoid false-positives, a successful MLAD system will take steps to help the user filter noise. Applying this model to cloud security, observing new users or devices, by definition, will generate patterns that are seen for the first time (a new IP address, a new application, a new account, etc. will appear). Good MLAD will learn source habits over time and flag anomalies only when, statistically, the event stream from a source, such as a user or device, is considered seasoned, or established enough.
More critically, MLAD must support a likelihood metric per event. Operators can display only the top N most unlikely/unusual events, sorted in descending order, while automatically filtering out any other event with a less than “one in a thousand,” or “one in a million” estimated probability to occur. Often these per-event likelihood metrics are based on the machine-learned statistical history of parameter values and their likelihood to appear together in context, for any source. It is up to the user to set the sensitivity thresholds to display what they want to see. This type of approach flags “rareness” and not “badness.”
Multi-dimensionality and generality: Successful MLAD platforms don’t rely on specific, hard-wired rules. Machine-learned anomalies are no longer unidimensional, such as “location-based,” “time-based,” etc. Instead, they are designed to detect anomalies in multiple, multi-dimensional spaces. You must look at every feature you can collect and that makes sense in every event and consider many features as a whole when calculating the likelihoods of each combination. An anomaly may be triggered due to one unusual value in a dimension, or a combination of multiple dimensions falling out of bounds. Features can be categorical or numeric, ordered or not, cyclical or not, monotonic or non-monotonic.
Robustness and ability to cope with missing data: Traditional batch machine learning clustering methods suffer from two critical issues:
- They break in the face of incomplete data, such as missing dimensions in some events.
- Due to the curse of dimensionality and the way distance metrics between multi-dimensional points are computed, they lose their effectiveness in high-dimensions (typically about five dimensions).
A good MLAD platform doesn’t rely on traditional batch clustering such as k-means. It is feature agnostic, dimension agnostic and can deal with missing or additional dimensions (features in an event) on the fly, as they appear.
Adaptability and self-tuning: Over time, even the most persistent habits tend to change. Users may switch to other applications, move to new geographical locations, etc. A platform that is based on machine learning adapts over time to new patterns and user habits. Old unusual patterns become the new norm if they persist for a long enough period. All conditional event probabilities keep updating over time.
Since organizations tend to be very different in the usage profiles, cloud app mix, event dimensions, and number of users, it’s important to keep a separate model for each organization and let it shift over time based on that organization’s change over time.
Future-proofing: An MLAD platform is agnostic to the semantics of input features. All it cares about is the statistical probability of each feature to occur in its specific context. We can add features (think about installing a new security camera, or any other new sensor) without code changes to the platform. The moment a new event source is introduced as a new input is the moment that you can detect anomalies in it. In the coming months, we plan to enrich our data with more features that will enable us to introduce additional data flows and models into the Netskope MLAD platform.
Personalization: Good MLAD studies each source separately, yet all in parallel. In Netskope’s MLAD, a source can be anything: a user, device, department, etc. In the real world, different sources tend to be very different in their normal behavior. In our experience, this fine-grained study of each source separately greatly improves signal-to-noise ratios.
Scalability: The algorithm we use can process tens of thousands of events per second per each tenant/model thread on standard hardware. We can run hundreds of such threads in parallel and horizontally scale as we grow. We can analyze the probability of any event in near constant time versus all prior historical events. The time to calculate the probability of any event is linear with the number of dimensions in the event. We detect anomalies as they come in, at a speed that is small constant multiplier over plain I/O of the same data.
User-friendliness: Each anomalous event is dissected and explained in-context using “smoking gun” evidence. For example, we may say, “This event is highly unusual (1 in 9.67 million likelihood) because, for this particular user, the source location is unusual, and the time of day is unusual, and this application has never been used before.” We do this while contrasting rare and unusual events with normal or common patterns. We don’t pass judgment on the maliciousness of an event; we only focus on likelihoods based on historical evidence. It is up to the user, given the information they have (and we don’t) to decide whether to take action on the information our anomaly-detection platform provides.
Dog food: As we were developing and testing MLAD at Netskope, we came across several interesting revelations. Users were downloading sensitive files from one app and then re-uploading those files to a separate app. After taking a closer look, we found that sensitive may have been been exfiltrated; file names included “Strategic Plan.pdf,” “passwords.txt” and “XYZ_litigation.docx.” This was one of the early indications that MLAD was kicking, and we were on the right track. Since then we’ve been discovering other unusual patterns that we had not anticipated before seeing MLAD in action.
“Solving security” is a tall order. Complex systems with many applications and users, and millions of possible access patterns can never be 100% secure. Our mission is to keep improving our tools. MLAD is one of these tools. We hope it will keep getting better and help our customers in their quest to keep their systems more secure.