Signal Processing Community and the Brazilian Federal Police Together in WIFS’11

by Jorge Lambert,  Anderson Rocha, et al  

Something really innovative has happened in Foz do Iguaçu, Brazil, where the third IEEE Workshop on Information Forensics and Security (WIFS) took place in November 2011. The 2011 WIFS, in a partnership with the Brazilian Association of Forensic Experts of the Federal Police (APCF), included in its program a series of oral presentations, posters, and demos, which represent an effort of the forensics experts to call the attention of academics to problems they often face.

The participation of the APCF was a first step toward a meeting (ICMedia) they are organizing in 2012, aiming at putting together experts, researchers, practitioners and solution developers to discuss forensics and security problems in Brazil in face of the upcoming 2014 World Cup and 2016 Summer Olympics. During 2011 WIFS, as a preliminary part of ICMedia, forensic experts from federal police and from all Brazilian states were encouraged to present real world Cases and Open Issues of special interest for daily tasks in forensic labs. This innovative format was only possible thanks to a joint work of forensic experts from the Brazilian Federal Police and researchers from the University of Campinas (Unicamp) represented by the 2011 WIFS Co-general chair Prof. Anderson Rocha.

The Cases and Open Issues included three 1-hour long oral sections with 12 open issues, four per section. In addition, there were nine posters and two tech-demos in a special poster section. The goal of each poster was to show an overview of some techniques used to solve real cases or to investigate crime traces provided that these techniques may be suitable for solutions in the Signal Processing community. The Open Issues presentations, on the other hand, focused on practical signal processing problems or proposals for which the solutions will improve forensics and police investigation activities worldwide and are, currently, considered difficult problems.

The Open Issues presented at 2011 WIFS

1. Chemical profile of illicit drugs: Brazilian Federal Police has evidence that many of its drug seizures are related to international drug trafficking. This evidence is based upon investigative work and information from police intelligence cooperation agreements worldwide, such as the DEA (USA). In several forensic analyses of cocaine and crack, chemical pattern of solvent residues can be used to draw traffic routes by identifying the laboratory where the drugs probably came from. There is high variability in the micro spectroscopy results, leading to a challenging pattern classification problem (source attribution).

2. Identification of marijuana and coca plantations from satellite imagery: Brazilian Federal Police has a growing issue with Cannabis crop identification. The illegal plantations take place in desert regions, northeast of Brazil (a.k.a. “Cannabis Polygon”). The Federal Police uses satellite imagery to identify growing plantations of coca and marijuana. There are several international initiatives to tackle similar problems. The issue is to automate the recognition as much as possible using signal processing and pattern recognition solutions.

3. Face recognition: computer aided methods for cephalometric landmarks marking : An outline of the method used for facial verification analysis, based on facial geometry and cephalometric fiducial points was presented. The questions pointed are Is it possible to measure facial proportions and correct perspective in CCTV images? What is the deviation tollerance? What about bad illumination conditions?

4. “DNA” DIAMOND PROJECT: Diamond gemstones source attribution: The major problem related to rough diamonds extracted illegally from conflict regions and smuggled is the attempt of obtaining Kimberley Process Certificate. There are ongoing works to determine the diamond source based on geological features observed by forensic experts. However, the method is very time-consuming, in a rate of about three diamond analyses per hour using a microscope in the field. The open issue here is related to a method based on digital images to determine the provenance of such diamonds.

5. Determining the writing anteriority in overlapping traces of handwritten documents obtained with different incident light wavelengths: Determining the priority in overlapped pen traces is one of the  problems in forensic analyses of handwritten documents. An effective technique based on Raman spectroscopy was outlined. However, many forensic labs do not have a Raman spectrometer. The open issue discussed how we can determine the writing anteriority in overlapped pen traces based on the analysis of digital images using different light wavelengths.

6. Content-based image retrieval, banknotes OCR and time-geographic money profile: Give money some traceability is effective to solve or prevent a lot of crimes. With this in mind, registering banknotes serial numbers might be useful. A 99,5% accurate OCR specific to the forensic task of reading and report Brazilian banknotes serial numbers was presented. In a given real case involving seizures of thousands of “random” R$ 100 banknotes, the histograms of such banknotes sets may suggest the existence of a time-geographic profile associated to large amounts of money, from which information can be retrieved. The open issues discussed were: (i) can we make the OCR solution more friendly and practical in order to enhance data acquisition from circulating money,  (ii) test the money profile conjecture and give it some statistical value?

7 & 8. Computer aided speaker segmentation and audio-based content retrieval: speaker dependent and independent approaches:  The poster and the presentation regarded automatic audio diasterization. The open issues discussed were: (i) can we have a practical solution for these tasks in noisy audio files? What prior information do we have to enter? (ii) Is it possible to search for child pornography in videos by analyzing the audio data in potentially porn videos?

9. Automatic detection of child porn on P2P networks:   The in-house developed software Espia Mule is currently used to detect child pornography files and find hundreds of criminals by analyzing network traffic in internet. The forensic demands in this case are solutions involving robots to probe P2P networks and automatically identify previously unknown child abuse material, estimate the total number of illicit files on these networks and total number of traders, possibly identifying the worst offenders.

10. Automatic detection of child pornographic images: Due to the need of proving criminal activities before seizing a lot of hardware or arresting suspects, law-enforcement officers must be able to detect child and adolescent pornography files among thousands of files stored on modern devices at suspects’ homes. The presented solution – Nu Detective – detects nudity performing pixel analysis and computational geometry, search for common pedophilia-related words and expressions and previously known hashes from black lists. The solution was designed to be used in suspect locations pointed by Espia Mule, so that an onsite preview can be performed by a forensic expert during an execution of a search warrant.. The open issue presented regards the inclusion of new and more sophisticated algorithms (e.g. with video analysis) in order to improve results without compromising time performance.

11. Content-based (semantic) image compression and CFTV super resolution:  In general, CCTVs are designed for surveillance only. So, it is often possible to get crime´s materiality and dynamics , but not authorship. This happens because compression algorithms are not content-oriented, ignoring objects of special interest to forensics, such as car plates, faces and guns. The open issue regards the proposal of a method to improve forensic application of CCTV images without significant costs by means of increasing the frame rate, applying super resolution algorithms in order to have better definition, identify regions of special interest (car plates, hands and maybe faces) and selectively compress these areas and background in order to concentrate losses in background and preserve quality on regions of forensic interest.

12. Calculating speed  and position of vehicles from low quality CCTV images: A real case of an aircraft accident was outlined. In this case, the speed of the plane was calculated from 320 x 240 images from CCTV cameras pointed to the runway at 5 fps. Using several reference points on the airplane, the speed could be calculated with less than 10% of estimated error. Obtaining speed estimation from low-quality CCTV images is a daily task in forensic labs worldwide. The open issue discussed how to obtain and select several reference points of a vehicle in such images and obtain speed with enhanced accuracy despite the low quality of CCTV system.