Writes the “About” section that “the goal of CRiTERIA project is a novel, comprehensive but feasible and human-rights sensitive risk and vulnerability analysis framework for border agencies, which backs a novel multi-perspective risk and vulnerability analysis methodology with multi-source, multi-lingual analysis technologies and tools for serving the complex indicators of the methodology and for making them accessible in a verifiable and understandable way.”
A link between migration and crime is established right from the start, and explicitly so: “Migration is increasingly framed as a security issue because immigrants are presumed to bring risks of terrorism, cross-border crime and illegal immigration (Dekkers et al., 2016).”
Developed functions are summarised in Newsletter #2:
1) a dashboard: “CRiTERIA’s visual analytics dashboard and the underlying open-access intelligence platform capture publicly available content from various government sources, NGOs, corporate websites, community platforms and social media channels.” Allegedly, this “ensures a comprehensive and nuanced analysis of the public debate, including real-time insights into migration-related issues, perceived problems, and emerging events such as public protests”. The dashboard would also allow its users to go beyond merely reactive measures, going as far as anticipating emerging risks: “Using the platform, stakeholders can respond proactively to emerging events and identify potential risks”
2) Event Detection in Media Reports
3) Migration Stance Shift (“by monitoring public sentiment towards refugees and migrants, officials can identify early warning signs of xenophobia or other negative attitudes”)
4) Video Event Recognition
5) Free-text Video Search
6) Revealing Hidden Networks
7) Stance Inference in Social Networks.
From a technological perspective, the aims of Working Package 5 (WP5) are of utmost interest, as they involve the collection and analysis of various types of data, including from social media websites, to produce predictions concerning migration dynamics:
“This WP develops and implements methods for the collection of multimodal (audio/video/text) data from online sources and their processing and analysis. The data (including associated metadata) coming out of the initial stages of analysis are used for identifying narratives, predicting trends, and ultimately supporting the detection of threat indicators from online and social media content. Specifically, WP5 develops methods for: the targeted collection of migration-related content from online sources, and the anonymization (or pseudo-anonymization) of the content where appropriate; the analysis of audio-visual content, to extract textual transcripts by means of (multilingual) automatic speech recognition (ASR), and visual media annotations in the form of concept and sentiment labels; the automatic detection and analysis of newly-evolving and influential narratives from multimodal content, building on top of the lower-level audio-visual analysis results (ASR, visual media annotation); the detection and prediction of trends, building explainable and reproducible deep learning techniques that not only make predictions from the analyzed content but also generate human-understandable explanations of them; consolidating (summarizing or fusing) the above results for detecting generic risk indicators from online, social media content.”
A lot of this needs clarification, especially concerning how to be consistent with the aims of WP4 (below).
According to the aims of WP6, data validation is also “semi-automatic.”
Among the stated aims of WP3 we can further, and interestingly, learn about the development of “composite indicators for risk analysis and vulnerability assessment which will enable border security to better tackle complex security problems”. Importantly, “one of the new categories of indicators is a sentiment indicator, which would enable practitioners to integrate beliefs and behaviours of risk communities/groups into their risk analysis.”
WP4 however highlights that “CRiTERIA is built on strong respect for the rule of law, fundamental rights in particular rights to privacy, data protection and expression. Moreover, CRiTERIA will be implemented taking in account the features of very complex phenomena and in alignment with social expectations of the involved actors. Gender and other intersectional differences will be carefully considered.”
Automated text analysis was adopted in a paper to investigate framings around Ukrainian migrants.
Also, the project worked on “Explainable AI” for visual classifiers in CRiTERIA, by developing : “T-TAME: Transformer-compatible Trainable Attention Mechanism for Explanations”. T-TAME is a post-hoc explainability method, i.e. it works with already-trained deep models and does not in any way degrade the accuracy of their decisions.”
Reinforcement learning techniques have also been adopted to enhance “News media profiling (reliability, political bias, factual reporting, etc.).”
Deliverable D4.1 details the “Relationship with other EC funded projects”.
Among them, we can find “H2020 projects such as: CITYCoP (Citizen Interaction Technologies Yield Community Policing); ARMOUR (A Radical Model of Resilience for Young Minds); MIRROR (Migration-Related Risks caused by misconceptions of Opportunities and Requirements); INSPECTr (Intelligence Network and Secure Platform for Evidence Correlation and Transfer) and PERCEPTIONS (Understanding the Impact of Narratives and Perceptions of Europe on Migration and Providing Practices, Tools and Guides for Practitioners)”.
These, writes the deliverable, “have assisted us in better understanding the legal and ethical problems that research with open source data can create for individuals. Lessons learned by these projects have been the starting point”.
The different components of the CRiTERIA system were tested and validated in 5 countries, ie. Estonia, Romania, Croatia, Malta and Sweden. “The results of each pilot will be closely monitored and fed to the risk analysis and technical team”, writes CRiTERIA’s website, “thus allowing the system to be improved from one pilot to the next”.
Activities “include the time planning, technical and logistical preparation of the labs and pilots, setup of pilot systems and analysis of their result. In the evaluation and validation of the CRiTERIA system, we will rely on benchmark tools for user satisfaction such as described alongside the KPIs earlier in the document.”
A focus was also put “on the development of a training module for border security staff aimed at familiarizing them with the novel risk analysis methodologies and tools.”
In terms of ethics assessment, Deliverable D4.2 consists of a “human rights implications checklist”, while D4.1 is a “Report of CRiTERIA Ethical Principles and Practices.”
Newsletter #3 (December 2023) writes: “Special focus is put on providing (semi-)automatic tools and methods for risk-related evidence validation and explanation and for identifying risk propagation and interlinking, thus, supporting decision processes in risk analysis in an innovative way. The methodology is developed in close collaboration with practitioners from border agencies, which will also validate the developed methods and technologies in piloting activities as well as with other stakeholders.”
A “Human Security Filter” is also described:
“The CRiTERIA consortium identified a critical gap in the humanitarian dimension of border security. Existing risk assessment methodology is commonly focussed on securing borders with little consideration of the vulnerabilities of migrants. Therefore, in CRiTERIA, we propose the creation of the so-called “Human Security Filter” which aims to reframe the risk assessment methodology bringing the human dimension of security into focus as a crucial part of risk considerations. Furthermore, it centralizes the humanitarian dimension of border protection, including the identification and addressing of migrant-specific vulnerabilities. As such, the Human Security Filter serves as a means of mainstreaming the human rights approach through the entire framework and guide border authorities in adequately assessing and addressing migrant vulnerabilities.”
A video on the Human Security Filter was released in February 2024. It is also described in greater detail — including its sub-indicators — here:
“The Human Security Filter (HSF) aims to measure a complex, multidimensional phenomenon, which cannot be measured directly. It aggregates several individual sub-indicators into a single composite measure, which is then seen as a reflection of the developments of the individual indicators. Currently, the HSF compromises 29 sub-indicators grouped into three categories (individual/embodied factors, situational factors, and structural factors) whereas the overall filter corresponds to the weighted averages of the sub-indicators. The weights are estimated in a so-called path regression model which includes both direct and indirect effects and provides estimates of the direction of the subsequent effects of the sub-indicators and the magnitude of the combined direct and indirect effects of the sub-indicators on the three main categories and the overall composite indicator. The sub-indicators and the three corresponding categories have been defined based on a comprehensive review of the available models in the literature (such as the IOM vulnerability framework and different European and national legal provisions on vulnerability) and expert consultations with 52 experts from various humanitarian organizations working in the field of migration.”
A study published by consortium members claim to be able to deduce what passive consumers of info on social media (“silent users”) actually think:
“current models usually rely on content production and overlook a vast majority of civically engaged users who passively consume information. These “silent users” can significantly impact the democratic process despite being less vocal. Accounting for the stances of this silent majority is critical to improving our reliance on SM to understand and measure social phenomena. Thus, this study proposes and evaluates a new approach for silent users’ stance prediction based on collaborative filtering and Graph Convolutional Networks, which exploits multiple relationships between users and topics. Furthermore, our method allows us to describe users with different stances and online behaviors. We demonstrate its validity using real-world datasets from two related political events. Specifically, we examine user attitudes leading to the Chilean constitutional referendums in 2020 and 2022 through extensive Twitter datasets.”
Several questions remain. Among them:
a) Early warnings of threats that might lead to violence are featured — including, as in a blogpost on rioting in Sweden, through disinformation. “Migration” here is therefore not only intended as referring to individuals, but to ideas as well (“ideas travel and lead to consequences”). How does the CRiTERIA project intend to automate these early warning signals/analyses?;
b) “Free-text video search is used in CRiTERIA to retrieve video content for any unpredictable query of the user, e.g. in the context of a study of migrant smuggling practices, retrieve all videos showing “A group of people walking in the woods”” How does this deep neural network work, more precisely?;
c) another blogpost provides an example of concrete “benefits” for the Malta Police, deriving from actual usage of the system. These include:
“1) Proactive Risk Mitigation: The CRiTERIA project equips the Malta Police Force with a novel risk analysis framework based on OSINT (Open Source Intelligence, ndr). By leveraging advanced analysis methodologies and technologies, law enforcement agencies gain the ability to identify potential risks and vulnerabilities before they escalate. This proactive approach allows for targeted interventions and preventive measures concerning immigration risks.
2) Data-Driven Decision Making: The CRiTERIA project integrates diverse open sources of data and employs advanced analysis techniques to generate actionable insights. By leveraging this data-driven approach, the police force gains a comprehensive understanding of complex indicators and can make informed decisions based on verifiable information. This could empower the Malta Police Force to allocate resources efficiently, optimize operational strategies, and respond effectively to evolving modus operandi.” These functions are both extremely sensitive and prone to automation bias: what has been done to ensure compliance with human rights?
