It is a known (among compiler developers) fact that checksums stored in the headers of Windows portable executable (PE) files are not actually checked by the operating system, with the exception of critical processes and libraries loaded into their address spaces (drivers, in particular). For this reason, it is not uncommon to come across a PE binary with an invalid checksum. In this post, a distribution of checksums collected from a typical Windows system is examined and compared to that for the checksums found in a mix of malware and benign modules from a known dataset. We look into the algorithm computing the checksum values in order to explain the results.

Investigates a curious phenomenon pertaining to decision trees, that of cross-entropy and accuracy metrics displaying coinciding (rather than opposite, as they normally do) trends.

Explains what exactly computing the expectation of a sample statistic and, in particular, an estimator entails.

Brings to the reader's attention a trivial property of joint expectations that many people use without giving it too much thought.

A widely-used approach to solving a non-convex optimization problem consists in deriving a series of convex problems (to which standard methods of convex optimizations apply) in such a way that their solutions converge to that of the original problem. In this post, two techniques for deriving such problem series are explained by example. We will also throw a cursory glance at what goes on "behind the scenes" as cvxpy converts an optimization program to the format understandable by ECOS.

Explains how to code a wheel installiation script in Python